<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Enrico De Santis on Medium]]></title>
        <description><![CDATA[Stories by Enrico De Santis on Medium]]></description>
        <link>https://medium.com/@enrico.desantis?source=rss-e9539446c6d8------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*totYmw-yADk_wQNtMhHJXA.jpeg</url>
            <title>Stories by Enrico De Santis on Medium</title>
            <link>https://medium.com/@enrico.desantis?source=rss-e9539446c6d8------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 13:10:44 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@enrico.desantis/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[When Knowledge Outpaces Understanding: the impact on thechnology]]></title>
            <link>https://medium.com/@enrico.desantis/the-when-knowledge-outpaces-understanding-the-impact-on-thechnology-df16787ff083?source=rss-e9539446c6d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/df16787ff083</guid>
            <category><![CDATA[fermats-last-theorem]]></category>
            <category><![CDATA[mathematics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[philosophy]]></category>
            <dc:creator><![CDATA[Enrico De Santis]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 09:36:41 GMT</pubDate>
            <atom:updated>2026-11-10T08:11:27.214Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Vp7ij5qlZJIsxBXfpO5b9A.avif" /></figure><h3><em>Artificial Intelligence, Mathematical Proof, and the Growing Autonomy of Technology</em></h3><p>In recent months, generative and agentic AI systems have begun producing proofs of mathematical problems that had resisted human efforts for decades. Some of these results have been formally verified or judged sound by independent mathematicians. Others, including OpenAI’s September 2026 announcement concerning the Navier–Stokes equations, have sparked debate over their validity and, more broadly, their scientific significance.</p><p>The Navier–Stokes case is particularly revealing. OpenAI announced a proof, accompanied by a formalization in the Lean proof assistant, showing that solutions to the three-dimensional Navier–Stokes equations can develop a singularity in finite time under a particular smooth external force. Although the construction addresses one of the alternatives included in the official formulation of the Millennium Prize Problem, several mathematicians have pointed out that it leaves unresolved what many regard as the more scientifically significant question of whether singularities can arise without external forcing. The Clay Mathematics Institute has acknowledged the announcement, while the formal process required to recognize a solution remains incomplete [1].</p><p>These developments have prompted an increasingly intense discussion within the mathematical community, beginning with the future of mathematical research itself. AI systems can now generate a growing volume of results, sometimes supported by proofs of extraordinary length and complexity. This raises difficult questions about <strong>verification, peer review, and the scientific community’s ability to absorb new knowledge</strong>. The <em>Leiden Declaration on Artificial Intelligence and Mathematics</em>, published in June 2026, has called attention to the need to preserve mathematical understanding, independent research, and reliable standards of verification [2].</p><p>Yet I believe a deeper question is emerging, one that concerns <strong>the very purpose of scientific inquiry and its relationship to technological progress</strong>.</p><h3>When a Proof Outruns Understanding</h3><p>Imagine an AI system producing a proof of a previously unknown mathematical theorem. The proof is formally correct and can be checked by an automated verification system, yet it is so long and intricate that no individual mathematician can meaningfully grasp it in its entirety.</p><p>We might have compelling formal grounds for accepting the theorem while remaining unable to explain fully why it holds, or to identify the broader mathematical principles underlying its proof.</p><p>What, then, would be the scientific value of such knowledge?</p><p>The proof would retain its logical and mathematical validity, provided its derivation had been correctly verified against the axioms and inference rules of the formal system. Formal verification, however, does not necessarily provide a conceptual explanation of the result. Nor does it automatically establish that the formal statement captures every aspect of the mathematical question we originally intended to address.</p><p>If we regard science as an endeavor devoted, among other things, to deepening our understanding of the world and the structures through which we describe it, we must acknowledge the possibility of a growing separation between <strong>the production of formally valid knowledge and its contribution to human understanding</strong>.</p><p>This distinction has a history that long predates artificial intelligence.</p><p>In his influential 1994 essay <em>On Proof and Progress in Mathematics</em>, William Thurston argued that mathematical progress involves far more than the accumulation of established theorems. It also depends on the development and transmission of understanding within the mathematical community [3].</p><p>More recently, Terence Tao has revisited this question in <em>Mathematics in the Age of AI</em>, examining the goals and values of mathematical research in a world where machines may become increasingly capable of solving problems whose solutions remain difficult for human mathematicians to interpret and assimilate [4].</p><p>The issue, as I see it, lies in the potential separation between <strong>formal correctness and scientific intelligibility</strong>. These properties have traditionally been closely intertwined, although there is no logical necessity for them to advance at the same pace.</p><p>A proof may be valid even when our conceptual understanding of it remains incomplete. Its formal correctness establishes a logical relationship within a mathematical framework, while understanding requires us to recognize the ideas, structures, and connections that give the result its broader significance.</p><p>Should the ability to generate formally verified results begin to grow much faster than the human capacity to interpret them, we could find ourselves facing an unfamiliar form of scientific abundance, one in which the expansion of mathematical knowledge increasingly exceeds our ability to make sense of it.</p><h3>From Scientific Understanding to Technological Utility</h3><p>Now consider another possibility.</p><p>Suppose that a theorem, despite its extraordinary complexity, makes it possible to design a new flight-control system capable of satisfying specific mathematical safety requirements under defined assumptions and operating conditions.</p><p>Once the mathematical model has been shown to represent the physical system adequately, and the necessary engineering validation has been carried out, the theorem could acquire enormous technological value.</p><p>We would then possess knowledge of considerable <strong>practical utility</strong>, even though our understanding of the mathematical reasoning supporting it remained incomplete.</p><p>An important distinction must be preserved here. The correctness of a mathematical proof does not, by itself, establish the reliability of a technological application. A theorem guarantees certain properties only within the assumptions and conditions under which it has been proved. Applying it to a physical system requires showing that those assumptions are satisfied and that the mathematical model adequately captures the relevant behavior of the real-world system.</p><p>In engineering, this means accounting for uncertainties, modeling limitations, disturbances, and operating conditions that may depart from idealized assumptions.</p><p>Yet this requirement makes the underlying question all the more interesting.</p><p>Once the necessary conditions have been established, <strong>the technological usefulness of a mathematical result may depend only weakly on our ability to understand the proof that establishes it</strong>.</p><p>This opens the door to a possibility that I find particularly significant.</p><p><strong>The validity of knowledge, our scientific understanding of it, and its technological usefulness may begin to develop along increasingly independent paths.</strong></p><p>The body of formally verified mathematical knowledge could expand far more rapidly than our ability to assimilate it. Meanwhile, the practical application of those results could continue to drive technological innovation.</p><p>Those technologies could then support further mathematical discoveries and new technological systems, giving rise to a cumulative process in which <strong>our capacity to act upon the world grows faster than our ability to understand the foundations of that capacity</strong>.</p><p>In such a scenario, the production of scientific knowledge might follow two partially divergent trajectories. One would remain oriented toward explanation, conceptual understanding, and the development of intelligible theories. The other would increasingly emphasize the generation and certification of mathematical results whose value could be assessed through the applications they enable.</p><p>These trajectories need not be mutually exclusive. Their relative importance, however, could change considerably as the production of mathematical results becomes increasingly automated.</p><p>The distinction would then acquire a new significance between the <em>epistemic aims of science</em>, understood as the pursuit of knowledge and understanding, and the <em>instrumental aims of technology</em>, directed toward achieving results according to particular criteria of utility.</p><p>The question becomes especially consequential if an increasing share of the mathematical foundations of technological development is eventually produced by artificial systems.</p><h3>The Long History of Problematic Proofs</h3><p>It is worth remembering that difficulties in validating mathematical proofs, including the possibility that serious errors may escape scrutiny for years, long predate artificial intelligence.</p><p>In 1879, Alfred Kempe published what was widely accepted as a proof of the four-color theorem. It was not until 1890 that Percy Heawood identified a fundamental flaw in the argument. Yet Kempe’s work introduced valuable mathematical techniques that survived the failure of the original proof and contributed to the eventual solution of the problem [5].</p><p>A different but equally instructive episode occurred in the work of Henri Poincaré. During the preparation of his celebrated study of the three-body problem, an error was discovered that required substantial revision. The resulting investigation helped reveal the extraordinary complexity of dynamical systems and laid important foundations for what would eventually become modern chaos theory [6].</p><p>More recently, Andrew Wiles announced a proof of Fermat’s Last Theorem in 1993. During the subsequent review, a serious gap emerged in the argument. Wiles ultimately resolved the difficulty in 1994, working with Richard Taylor, and the completed proof was published in 1995 [7].</p><p>These episodes reveal a remarkable feature of mathematical progress. <strong>An argument may fail as a proof while still containing ideas capable of transforming an entire field.</strong> The discovery of an error can even lead to mathematical insights that extend far beyond the original problem.</p><p>There is another historical dimension to consider.</p><p>Mathematics has also produced proofs whose correctness was difficult to establish because of their sheer complexity or dependence on extensive computation.</p><p>The 1976 proof of the four-color theorem by Kenneth Appel and Wolfgang Haken generated considerable controversy because it relied on computer calculations that could not be reproduced through conventional, line-by-line human inspection.</p><p>Similarly, Thomas Hales’s 1998 proof of the Kepler conjecture, concerning the densest possible packing of equal spheres, proved so demanding to verify that it eventually led to the <em>Flyspeck Project</em>, an ambitious effort to formalize the entire argument using proof assistants. The project was completed in 2014, with the formal proof subsequently published in 2017 [8].</p><p>Artificial intelligence is therefore entering a mathematical landscape in which questions about the reliability, intelligibility, and verification of proofs have existed for centuries.</p><p>What is changing is the potential <strong>scale, speed, and autonomy of mathematical production</strong>.</p><p>AI systems may eventually generate results at a pace that far exceeds the scientific community’s capacity to review, interpret, and integrate them into a coherent body of mathematical understanding.</p><p>There is a further implication worth considering.</p><p>An incorrect AI-generated proof could still contain fruitful mathematical insights, just as flawed human arguments have done throughout the history of mathematics. It might suggest a previously unknown technique, introduce a useful construction, or reveal unexpected relationships between mathematical structures.</p><p>The <strong>heuristic value</strong> of an argument may survive its formal invalidity, even though applications requiring rigorous guarantees must ultimately rest on correctly established properties.</p><p>This possibility deserves particular attention as AI-generated mathematical research continues to expand.</p><p>A proliferation of incorrect proofs could undoubtedly overwhelm existing review processes and undermine confidence in the mathematical literature. At the same time, some of those unsuccessful attempts might contain ideas worth recovering, refining, and validating.</p><p>The growing ability of AI systems to explore mathematical possibilities could therefore change the relationship between error and discovery.</p><p>A far larger universe of candidate proofs, including failed attempts, might become available for examination. The scientific challenge would increasingly involve recognizing which of those attempts contain structures, methods, or insights capable of supporting genuine advances in knowledge.</p><p>In this respect, the ability to distinguish mathematical insight from formal error could become an increasingly demanding part of scientific research.</p><h3>Toward a New Autonomy of Scientific and Technological Development</h3><p>All of this leads me to a broader question.</p><p>For a long time, we have tended to imagine technological progress as the outcome of increasingly sophisticated scientific understanding.</p><p>Of course, the history of technology offers many examples of practical devices developed before their underlying scientific principles were fully understood. The steam engine, for instance, preceded the mature theoretical framework of thermodynamics.</p><p>Artificial intelligence, however, introduces a further possibility.</p><p>We may be approaching a situation in which <strong>mathematical results are generated, formally verified, and put to technological use long before human beings are able to understand them in depth</strong>.</p><p>The distinctive feature of this development would lie in the ability of technological systems to participate in generating the mathematical foundations of their own further evolution.</p><p>AI systems capable of producing proofs could contribute to the development of new algorithms and technologies, which might subsequently support further discoveries and technological advances.</p><p>Human researchers would continue to establish goals, define constraints, evaluate results, and oversee these processes. Yet such supervision would not necessarily entail a complete understanding of every mathematical step on which the resulting technologies depend.</p><p>A new relationship between knowledge and application could begin to emerge.</p><p><strong>Scientific understanding might increasingly follow the technological use of a mathematical result, rather than precede it.</strong></p><p>Understanding would then become, in some cases, the subject of a subsequent scientific investigation, undertaken after the result had already been certified and put to practical use.</p><p>The implications extend well beyond mathematics.</p><p>If the scientific and technological apparatus gradually acquires the ability to generate the theoretical foundations of its own development, what becomes of the relationship between the pursuit of knowledge and the pursuit of technological capability?</p><p>This question resonates with the philosophical work of Emanuele Severino, particularly his reflections on the growing autonomy of the technological apparatus and its tendency to evolve from a means of achieving externally defined goals into a system increasingly capable of shaping those goals itself [9].</p><p>From this perspective, artificial intelligence may introduce a new dimension to a process already examined in the philosophy of technology.</p><p>As technological systems become capable of producing mathematical knowledge that supports their own further development, the relationship between scientific understanding and technological power may undergo a profound transformation.</p><p>The technological apparatus could increasingly acquire the means to expand its capabilities through knowledge that it helps produce, while human understanding of that knowledge develops at a slower pace.</p><p>A process of this kind would raise questions that extend beyond the reliability of individual mathematical proofs. It would concern the direction of scientific inquiry itself and the possibility that technological development might progressively influence which forms of knowledge are produced, pursued, and valued.</p><p>This remains a largely hypothetical scenario. The current achievements of AI in mathematics are remarkable, yet they provide no grounds for concluding that mathematical understanding will eventually become unnecessary or that technological development could proceed indefinitely without meaningful human oversight.</p><p>They do, however, provide compelling reasons to examine the possibility that the <strong>production, verification, and application of knowledge may become increasingly separated from its conceptual assimilation by human beings</strong>.</p><p>Perhaps one of the deepest philosophical questions raised by artificial intelligence concerns precisely this emerging divergence between <strong>what we can prove, what we can understand, and what we can build</strong>.</p><p>Ultimately, we may be entering an era in which <strong>the technological power of knowledge continues to grow even as our scientific understanding of its foundations remains incomplete</strong>.</p><p>If this possibility becomes a defining feature of scientific progress, we will have to reconsider what it means to advance human knowledge in the first place, and what place human understanding should occupy within a scientific enterprise increasingly capable of extending its own boundaries.</p><h3>References</h3><p>[1] OpenAI (2026). <em>On the Navier–Stokes Millennium Prize Problem</em>. September 8. <a href="https://openai.com/index/navier-stokes-solution/">OpenAI publication</a>. Clay Mathematics Institute (2026). <em>Navier-Stokes Announcement</em>. September 11. <a href="https://www.claymath.org/news/navier-stokes-announcement/">Official statement</a>.</p><p>[2] Alper, J., et al. (2026). <em>Leiden Declaration on Artificial Intelligence and Mathematics</em>. <a href="https://doi.org/10.5281/zenodo.20302944">https://doi.org/10.5281/zenodo.20302944</a>.</p><p>[3] Thurston, W. P. (1994). <em>On Proof and Progress in Mathematics</em>. <strong>Bulletin of the American Mathematical Society</strong>, 30(2), 161–177. <a href="https://doi.org/10.1090/S0273-0979-1994-00502-6">https://doi.org/10.1090/S0273-0979-1994-00502-6</a>.</p><p>[4] Tao, T. (2026). <em>Mathematics in the Age of AI</em>. Essay based on a lecture delivered at the International Congress of Mathematicians 2026. <a href="https://arxiv.org/abs/2608.16753">arXiv:2608.16753</a>.</p><p>[5] Wilson, R. (2016). <em>Wolfgang Haken and the Four-Color Problem</em>. Celebratio Mathematica. <a href="https://celebratio.org/Haken_W/article/793/">Historical account</a>.</p><p>[6] Barrow-Green, J. (1997). <em>Poincaré and the Three Body Problem</em>. History of Mathematics, Vol. 11. American Mathematical Society and London Mathematical Society.</p><p>[7] Wiles, A. (1995). <em>Modular Elliptic Curves and Fermat’s Last Theorem</em>. <strong>Annals of Mathematics</strong>, 141(3), 443–551. <a href="https://doi.org/10.2307/2118559">https://doi.org/10.2307/2118559</a>. Taylor, R., &amp; Wiles, A. (1995). <em>Ring-Theoretic Properties of Certain Hecke Algebras</em>. <strong>Annals of Mathematics</strong>, 141(3), 553–572. <a href="https://doi.org/10.2307/2118560">https://doi.org/10.2307/2118560</a>.</p><p>[8] Hales, T., et al. (2017). <em>A Formal Proof of the Kepler Conjecture</em>. <strong>Forum of Mathematics, Pi</strong>, 5, e2. <a href="https://doi.org/10.1017/fmp.2017.1">https://doi.org/10.1017/fmp.2017.1</a>.</p><p>[9] Severino, E. (1998). <em>Il destino della tecnica</em>. Milan, Rizzoli.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=df16787ff083" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When a Machine Seems to Think: Douglas Hofstadter and Noosemia]]></title>
            <link>https://medium.com/@enrico.desantis/when-a-machine-seems-to-think-douglas-hofstadter-and-noosemia-0e95d8edb4df?source=rss-e9539446c6d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/0e95d8edb4df</guid>
            <category><![CDATA[noosemia]]></category>
            <category><![CDATA[philosophy-of-mind]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cognitive-science]]></category>
            <dc:creator><![CDATA[Enrico De Santis]]></dc:creator>
            <pubDate>Fri, 16 Oct 2026 17:38:31 GMT</pubDate>
            <atom:updated>2026-10-16T17:38:31.915Z</atom:updated>
            <content:encoded><![CDATA[<p><em>From the ELIZA Effect to Large Language Models: The Shifting Boundary Between Understanding, Mind Attribution, and Disenchantment</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W4O7_BfB7RM9Y2IT9I21lg.png" /></figure><blockquote>“It is nothing but the verbal behaviour of such systems that makes at least some people convinced that behind the scenes there reside meanings, concepts, ideas, beliefs, thinking, sentience and even consciousness.”— Douglas R. Hofstadter, “Is there an ‘I’ in AI?”, Philosophical Transactions of the Royal Society A, 384, 20240527 (2026), p. 6. <a href="https://doi.org/10.1098/rsta.2024.0527?utm_source=chatgpt.com">https://doi.org/10.1098/rsta.2024.0527</a></blockquote><h3>The Puzzle of a Convincing Answer</h3><p>What does it take for us to recognize a sign of intelligence? This question has occupied cognitive science for decades, acquiring particular significance when the behavior we observe comes from a machine. A relevant answer, an unexpected analogy, or a convincing solution to a problem may suggest an underlying understanding that we cannot directly observe. This is a familiar feature of human relationships, in which the words and actions of others provide our primary access to their mental lives. Generative artificial intelligence has brought a new dimension to this experience, introducing artificial systems capable of increasingly sophisticated linguistic behavior and renewing some of the deepest questions in the philosophy of mind.</p><p>Douglas Hofstadter is among the scholars who have contributed most significantly to exploring this territory. His <em>Gödel, Escher, Bach: An Eternal Golden Braid</em>, published in 1979 and awarded the Pulitzer Prize in 1980, brings together mathematical logic, music, art, and cognitive science in an exploration of self-reference and the possible emergence of a self. His intellectual journey has taken him from the early development of symbolic artificial intelligence through the study of analogy and conceptual representation to the more recent advances in deep neural networks.</p><p>Hofstadter recounts an episode involving a deceptively simple task. He asks GPT-4 — the model introduced by OpenAI on March 14, 2023, and later retired from the ChatGPT interface on April 30, 2025 — to explain how to build a stable stack using a book, nine eggs, a laptop, a bottle, and a nail (OpenAI, 2023, 2025). Solving the problem requires an appreciation of the objects’ physical properties and dimensions, the ability to imagine how they might be arranged in space, and some understanding of what would allow the resulting structure to remain balanced.</p><p>GPT-4 proposes a systematic solution. It suggests using the book as a base, arranging the nine eggs on top in a three-by-three square, placing the closed laptop over the eggs, and then adding the bottle and the nail, with the latter standing vertically on the bottle cap. The actual stability of this arrangement is open to question, but Hofstadter is struck by the model’s ability to produce a description that seems to account for the three-dimensional relationships among real-world objects.</p><p>In his essay <em>Is there an ‘I’ in AI?</em>, published on May 14, 2026, in <em>Philosophical Transactions of the Royal Society A</em>, he describes his reaction:</p><blockquote><em>This description of how to make a pile out of the given objects constituted, at least for me, a stunning match of words and phrases with the way the real world actually works — an intimate alignment of words with three-dimensional reality. Where does that come from, if not from understanding?</em></blockquote><p>(Hofstadter, 2026, p. 1).</p><p>His closing question makes the episode particularly revealing. Hofstadter is entertaining the possibility that the model’s linguistic performance reflects some form of understanding. Its words, apparently aligned with the physical world, seem to point toward something we would ordinarily associate with thought.</p><p>The qualification <em>at least for me</em> deserves particular attention. Hofstadter acknowledges that his impression belongs to an act of interpretation: something generated by a machine has acquired, in his experience, the significance of a possible sign of cognitive activity.</p><p>In work developed with Antonello Rizzi, we introduced the term <strong>noosemia</strong> to describe precisely this dimension of experience, in which what an entity expresses or produces is recognized as a sign of possible mental activity. In its most concise formulation, noosemia is <em>the experience of a sign as a sign of mind</em> (De Santis &amp; Rizzi, 2025; De Santis, 2026).</p><p>Hofstadter’s account becomes even more interesting as the episode unfolds. His initial impression is soon challenged by an error in the model’s answer. What had seemed to reveal understanding becomes a reason for doubt, bringing into view the reversible nature of mind attribution.</p><h3>From ELIZA to Noosemia</h3><p>This episode has particular historical significance because Hofstadter had addressed the problem of attributing understanding to computers long before large language models appeared.</p><p>In <em>Fluid Concepts and Creative Analogies: Computer Models of the Fundamental Mechanisms of Thought</em>, published in 1995, he included a preface titled <em>The Ineradicable Eliza Effect and Its Dangers</em> (pp. 155–168). There, Hofstadter examined the risks of attributing to artificial intelligence programs a degree of understanding that their actual mechanisms could not justify. His discussion belonged to an established tradition of late-twentieth-century cognitive research, when artificial intelligence sought to reproduce abilities regarded as distinctively human through symbolic processing, knowledge representation, and analogy-making.</p><p>The reference to ELIZA takes us back to one of the earliest and best-known conversational programs. Developed by Joseph Weizenbaum at MIT and described in a 1966 article in <em>Communications of the ACM</em>, ELIZA used pattern-matching and transformation rules to sustain simple conversations. In its best-known configuration, called DOCTOR, the program imitated the conversational style of a Rogerian psychotherapist, drawing on parts of a user’s statements to formulate further questions and keep the exchange going (Weizenbaum, 1966).</p><p>Despite the relative simplicity of its mechanisms, some users were inclined to experience the interaction as though the program understood what they were saying. The conversational format created an impression of attentiveness and relevance that could arise without any knowledge of the underlying procedures. The phenomenon, which became known as the <em>ELIZA effect</em>, revealed how readily people may interpret a machine’s linguistic behavior through categories normally associated with minds.</p><p>Hofstadter recognized a profound epistemological problem in this tendency. The richness of meaning attributed by the interlocutor could far exceed what the program’s operations warranted, particularly when a convincing exchange made its mechanical character less apparent.</p><p>Decades later, the same problem arises in a radically different technological landscape. The introduction of the Transformer architecture, described by Vaswani and colleagues in 2017, helped drive the development of models capable of learning distributed representations from vast amounts of data and generating responses adapted to complex linguistic contexts. Contemporary generative systems can perform tasks involving specialized knowledge, analogical reasoning, semantic transformations, and sophisticated problem-solving, although their limitations depend on the particular model and the conditions under which it is used.</p><p>The historical connection to ELIZA lies in the way human beings interpret artificial behavior. The technological differences, meanwhile, require us to evaluate each system according to its actual capabilities and to avoid projecting the limitations of earlier generations onto their successors.</p><p>It is against this background that we proposed the concept of noosemia, formally introduced in 2025 to examine the cognitive and phenomenological dimensions of intentionality attribution in human interaction with generative AI (De Santis &amp; Rizzi, 2025).</p><p>The term is a modern scholarly coinage formed from two families of Greek-derived words. The element <em>noo-</em> refers to <em>nóos/noûs</em> (νόος/νοῦς), associated with mind and intellect, while <em>-semia</em> belongs to the linguistic family of signs and signification, related to <em>sêma</em> (σῆμα), <em>sēmeîon</em> (σημεῖον), and <em>sēmaínein</em> (σημαίνειν). Its conceptual roots thus connect the sign with <em>nóēsis</em> (νόησις), or intellectual activity. The word is a contemporary formation intended to identify a specific phenomenon of interpretive experience (De Santis, 2026).</p><p>In the semiotic-evidential formulation developed subsequently, <strong>noosemia is the relational and semiotic phenomenon through which an observable manifestation acquires, in the experience and interpretation of a subject within a given context, the status of an evidential sign of noetic activity attributed to the source of that manifestation</strong>.</p><p>We can approach this definition by considering how the significance of an answer may change during an interaction. When a machine grasps an implication that was never explicitly stated, corrects its own solution, or proposes an unexpected yet relevant connection, the person interacting with it may find themselves recognizing something that appears to arise from an understanding of the problem. The response takes on a significance that extends beyond the mere fact of its production.</p><p>The <strong>noosemic effect</strong> refers to the way this dynamic manifests itself in the observer’s experience. Recognition may be accompanied by surprise or engagement and, over a prolonged interaction, may contribute to a growing sense of familiarity with the system. At its core, however, the phenomenon concerns the meaning attributed to what is observed; whether that attribution is justified remains a separate question.</p><p>This approach also clarifies an important distinction from the traditional interpretation of the ELIZA effect. The latter is generally associated with attributing to a system a degree of understanding that exceeds its actual capabilities. Noosemia concerns the process through which such an attribution arises, regardless of whether it later proves justified, disproportionate, or impossible to assess conclusively.</p><p>This distinction allows us to approach contemporary generative systems without assuming in advance that every impression of understanding must be illusory. The question becomes how something produced by a system comes to be experienced as a sign of mind, and what may strengthen, alter, or weaken that recognition.</p><h3>When Behavior Becomes a Sign of Understanding</h3><p>In <em>Is there an ‘I’ in AI?</em>, Hofstadter directly addresses the relationship between language and thought. His argument begins with the idea that the words produced by a system may acquire meaning insofar as they correspond to the way the world works. When a description appropriately captures relationships in reality, he finds it difficult to deny that it reflects some degree of understanding.</p><p>The essay’s title reveals the connection between this position and the problem of self-consciousness. For Hofstadter, the question of artificial minds ultimately leads to the conditions under which a system might develop a model of itself.</p><p>The publication history of the essay deserves some clarification. Although the article appeared in May 2026, an earlier manuscript signed by Hofstadter and dated April 2023 already contains the core argument and several of the episodes discussed here. The version published by the Royal Society was received on January 21, 2025, and accepted on June 19 of the same year. The reflections presented in the article should therefore not be understood as a position that emerged only in 2026.</p><p>In a central passage, Hofstadter observes:</p><blockquote><em>It is nothing but the verbal behaviour of such systems that makes at least some people convinced that behind the scenes there reside meanings, concepts, ideas, beliefs, thinking, sentience and even consciousness.</em></blockquote><p>(Hofstadter, 2026, p. 6).</p><p>This observation closely resembles the central idea of noosemia. Hofstadter describes how linguistic behavior can lead people to attribute mental properties to the system producing it. What the machine says becomes a basis for inferring something that cannot be verified through direct access to any internal experience the system might possess.</p><p>The case is especially interesting because Hofstadter is thoroughly familiar with the computational nature of machines. His impression arises from a sustained reflection on the limitations of artificial architectures and the difficulty of determining what should count as genuinely cognitive. Technical knowledge informs his assessment, yet the model’s performance can still prompt him to reconsider whether understanding might be present.</p><p>In the same essay, Hofstadter proposes a condition under which he would recognize thought in a language model:</p><blockquote><em>I would say that if a large language model successfully and consistently combines ideas in ways that make sense, then I would have to grant that it is thinking.</em></blockquote><p>(Hofstadter, 2026, p. 4).</p><p>Here the discussion reaches a further philosophical level. Hofstadter considers the ability to combine ideas meaningfully and consistently as a possible reason to attribute thought to an artificial system. Noosemia allows us to examine how that same coherence comes to be experienced as a sign of mental activity, while leaving the justification of the attribution open to further inquiry.</p><p>This distinction becomes essential when a successful performance is followed by an error that seems incompatible with the understanding previously attributed to the system. That is precisely what happens in the stacking experiment, where Hofstadter’s initial astonishment soon gives way to a much deeper doubt.</p><h3>When the Impression of Understanding Falters</h3><p>Let us return to the stack of objects. GPT-4’s answer contains a peculiar claim: the model suggests that the nail’s sharp tip, pointing upward, would prevent it from rolling or sliding off the bottle. The explanation makes little physical sense, disrupting the impression of coherence created by the preceding description.</p><p>Hofstadter responds by asking:</p><blockquote><em>How can a vertical nail’s sharp tip (pointing upwards) prevent it from rolling or sliding? That doubt on my part makes me wonder if GPT-4 genuinely understands (in the sense of having something akin to visual imagery) what any of the objects in the pile are really like, physically.</em></blockquote><p>(Hofstadter, 2026, p. 2).</p><p>The change is significant. The very answer that had initially suggested understanding is reconsidered in light of a flaw that changes how Hofstadter interprets the entire performance. The error does not erase what the model had described correctly, but it weakens the inference he had begun to draw about its grasp of the physical world.</p><p>Melanie Mitchell, an artificial intelligence researcher and longtime interlocutor of Hofstadter, makes the problem even more apparent. After reading an earlier version of his essay, she asks ChatGPT, using GPT-4, to build a stable stack from a toothpick, a bowl of pudding, a marshmallow, and a full glass of water. The system suggests inserting the toothpick into the pudding, placing the marshmallow on its tip, and finally balancing the full glass on the marshmallow. The proposed arrangement reveals a striking failure to account for the physical stability of the objects.</p><p>The episode prompts Hofstadter to reconsider whether language models possess concepts at all. His reaction includes a particularly revealing statement:</p><blockquote><em>Indeed, gross blunders like this are so silly — so wildly clueless — that they reawaken, in my mind, great doubt as to whether LLMs possess concepts at all. I repeat: having this doubt is a huge relief to me.</em></blockquote><p>(Hofstadter, 2026, p. 2).</p><p>His expression of relief offers an important insight into the emotional dimension of the experience. Hofstadter had initially recognized something resembling understanding in the model’s performance; the discovery of a glaring error restores a sense of distance between its behavior and the capacities he associates with thought.</p><p>Within the framework proposed by De Santis and Rizzi (2025), this movement can be understood through <strong>a-noosemia</strong>, the phenomenological withdrawal of a previously established noetic attribution. When a subsequent response contradicts the expectations that supported the initial recognition, the behavior may lose some of the significance it had acquired as evidence of mental activity.</p><p>Hofstadter’s account is particularly valuable because it reveals an interpretive sequence through reactions he explicitly reports. Initial astonishment, subsequent doubt, and relief belong to the same experience, showing how the recognition of a possible artificial mind can be reconsidered as an interaction unfolds. The suspension of that recognition may also be temporary, since further coherent performances could support it once again.</p><p>Other interactions may produce little or no sense of mental presence from the outset. Hofstadter describes, for example, his experience with an automated telephone system that uses first-person expressions and claims to understand complete sentences. Despite its seemingly appropriate language, he senses no significant mental presence behind its responses:</p><blockquote><em>What kind of ‘I’-ness is there behind the scenes of such a telephone rotbot? Precious little. In fact, zero — or at least damn close to zero. Thermostat-level ‘I’-ness, let’s say, or maybe cockroach-level.</em></blockquote><p>(Hofstadter, 2026, p. 8).</p><p>This second example helps clarify an important distinction. In the stacking experiment, the recognition of possible understanding is weakened by a subsequent response; in the telephone system example, Hofstadter describes an interaction in which mind attribution barely arises at all. These experiences have different temporal structures, a distinction that matters when separating a-noosemic withdrawal from the initial absence of attribution.</p><p>The dynamic becomes even more interesting when Hofstadter begins to question how he himself evaluates mistakes made by human beings and machines.</p><h3>The Observer’s Role in Recognizing a Mind</h3><p>Hofstadter recounts reading a poem generated by GPT-4 that presented Euclid’s proof of the infinitude of prime numbers. Certain inaccuracies had led him to doubt whether the model genuinely understood the mathematical argument. Later, his nephew Nathaniel, who holds a doctorate in mathematics, sends him a poem on the same subject containing a more serious mistake: it omits the crucial step of adding one to the product of the prime numbers under consideration.</p><p>Yet Hofstadter is forgiving of his nephew’s error. Knowing Nathaniel’s mathematical competence, he interprets the omission as an oversight entirely compatible with genuine understanding of the proof. Reflecting on his different reactions, he acknowledges:</p><blockquote><em>But I didn’t jump all over Nathaniel, for I knew well that he knew exactly what he was doing but had just accidentally overlooked something small (like me searching for my glasses while wearing them), whereas I took GPT-4’s much tinier math error (and arguably it wasn’t even an error at all!) as evidence that it didn’t know what it was talking about. That constituted an undeniable double standard on my part.</em></blockquote><p>(Hofstadter, 2026, pp. 3–4).</p><p>His recognition of this asymmetry is theoretically significant. Hofstadter acknowledges that we interpret a performance partly in light of what we already know about the person or system producing it. When someone whose competence we trust makes a mistake, we are willing to treat it as an isolated incident. Faced with an artificial system, however, we may assign a much broader meaning to an error of a similar kind.</p><p>Noosemia offers a way to understand this difference through the relational character of mind attribution. An answer comes to be perceived as a sign of mind within an interpretive history, in which expectations about the interlocutor shape the significance of subsequent behavior. The same performance may therefore elicit different judgments depending on who observes it and what they believe they know about its origin.</p><p>Hofstadter extends his reflection to the way the linguistic abilities of machines can change people’s beliefs about whether computers might think or be conscious:</p><blockquote><em>When computers seem to have rich language and rich meanings, many people change their tune very quickly (without even noticing that things are insidiously shifting inside their heads) about whether computers could ever think or be conscious.</em></blockquote><p>(Hofstadter, 2026, p. 5).</p><p>The system’s behavior thus participates in reshaping the observer’s interpretive framework. Such changes may develop gradually through successive interactions and become apparent only when a new performance challenges assumptions about what a machine can do.</p><p>In this connection, Hofstadter recalls experiments he conducted with GPT-3, the model described in OpenAI’s research in May 2020. It had produced answers that were sometimes comically inadequate when faced with questions designed to test common sense. In his essay, he notes how much performance had changed with GPT-4, introduced in March 2023. He also discusses LaMDA, the conversational technology announced by Google on May 18, 2021, whose ability to explain certain jokes had particularly impressed him (OpenAI, 2020, 2023; Collins &amp; Ghahramani, 2021).</p><p>GPT-3 and LaMDA refer to research projects and model families for which there is no single retirement date applicable to every implementation. For the original GPT-4, by contrast, several milestones can be distinguished. Following its introduction on March 14, 2023, it was retired from the ChatGPT interface on April 30, 2025. The retirement of specific legacy models and identifiers available through the API, including gpt-4 and gpt-4-0613, was scheduled for October 23, 2026. At the time of writing, that deadline is still in the future. Hofstadter&#39;s texts do not provide enough information to identify the exact model snapshot used in each experiment (OpenAI, 2023, 2025, 2026).</p><p>This chronological precision matters for reasons that extend beyond reconstructing technical history. Assessments of artificial capabilities always belong to a particular stage of technological development. A limitation that seems characteristic of one model may be overcome by a subsequent generation, while more advanced capabilities may introduce new failures or unexpected behavior. As these systems evolve, our conclusions must remain anchored to the conditions under which they were reached.</p><p>Such caution is especially important when discussing broad questions such as understanding or artificial consciousness. A limitation observed in a particular system provides valuable evidence about that system, but it cannot establish once and for all what artificial systems will or will not be able to achieve. The recent history of AI shows how rapidly the capabilities underlying our judgments can change.</p><p>The noosemic significance of artificial behavior also changes as systems improve and observers adjust their expectations. An answer that initially inspires wonder may later become familiar; an unexpected performance may reopen questions that had seemed settled. The relationship between observer and machine therefore develops through an evolving history of experience, interpretation, and revision.</p><h3>An Artificial “I” Speaking in Hofstadter’s Name</h3><p>A second episode, recounted by Hofstadter in <em>Gödel, Escher, Bach, and AI</em>, published in <em>The Atlantic</em> on July 8, 2023, allows us to approach mind attribution from another perspective.</p><p>Sami Al-Suwailem, a reader of <em>Gödel, Escher, Bach</em>, wanted to publish a brief account of how the book came into being. Rather than asking its author to write the piece, he used GPT-4 to generate a first-person text, as though Hofstadter himself had written it. The request referred to the twentieth-anniversary edition of the book, which contains an autobiographical preface, but there was no evidence that the model had actually consulted that material.</p><p>Hofstadter describes the episode as follows:</p><blockquote><em>The AI system obliged Al-Suwailem by instantly producing a statement written in the first person (thus impersonating me, the book’s author), which Al-Suwailem sent to me via email, hoping to get my approval for posting it online.</em></blockquote><p>(Hofstadter, 2023).</p><p>The generated text touched on many themes genuinely present in the book. It referred to self-reference and recursion, the relationships between logic, art, and music, and the connection between intelligence and consciousness. Its solemn prose seemed broadly appropriate for an intellectual autobiography.</p><p>Hofstadter, however, recognized a succession of generalities that did not correspond to the actual history of the book’s creation, his intentions, or his way of writing. In his reply to Al-Suwailem, subsequently reproduced in the article, he observed:</p><blockquote><em>Although someone who was unfamiliar with my writing might take this saccharine mixture of pomposity and humility as genuine, to me it is so far from my real voice and so far from GEB’s real story that it is ludicrous.</em></blockquote><p>(Hofstadter, 2023).</p><p>The difference between a reader’s possible impression and the author’s own experience is particularly revealing. The text might appear plausible to someone familiar only with the book’s general themes, whereas Hofstadter’s knowledge of his personal history and the decisions that shaped the writing process makes the discrepancies immediately apparent.</p><p>To explain his reaction, Hofstadter reconstructs the actual origins of <em>GEB</em>. He recalls reading <em>Gödel’s Proof</em> by Ernest Nagel and James R. Newman in his youth, a book that introduced him to the fascination of the incompleteness theorem, followed by Howard DeLong’s <em>A Profile of Mathematical Logic</em>. He describes a long letter written in 1972 to his friend Robert Boeninger, in which he began exploring the relationship between Gödel’s ideas and the mystery of the self, and an initial manuscript completed in 1973, which still contained no illustrations, references to Bach, or literary dialogues.</p><p>The book’s eventual form took shape through a longer creative process. The dialogues between Achilles and the Tortoise grew out of Lewis Carroll’s influence, while the decision to organize them around musical counterpoint gradually brought Bach into the project. Escher’s images, which had already played a role in Hofstadter’s thinking, became an explicit part of the book at a later stage, helping to shape its distinctive literary architecture.</p><p>These episodes belong to the concrete history of an intellectual and autobiographical undertaking, shaped by particular readings, insights, and successive transformations. The narrative generated by GPT-4 offered an abstract account of that history, assembled from relevant themes but lacking the specificity Hofstadter recognized as his own.</p><p>His conclusion is emphatic:</p><blockquote><em>In short, the piece that GPT-4 composed using the pronoun “I” has zero authenticity, it has no resemblance to my manner of expressing myself, and the artificiality of its creation runs against all the pillars of my lifelong belief system.</em></blockquote><p>(Hofstadter, 2023).</p><p>The problem of authenticity adds another dimension to the relationship between signs and minds. An artificial system can produce first-person statements, apparently reconstruct intentions and memories, and imitate an individual voice. The meaning of such expressions also depends on the relationship between the text and the person it purports to represent.</p><p>In Hofstadter’s account, GPT-4 generated the text, while the identity it assumed was that of the scholar himself. His rejection therefore concerns a particular form of artificial impersonation, in which the appearance of an authentic voice fails to correspond to the author’s personal and intellectual history.</p><p>The episode is relevant to noosemia because it shows how knowledge of the purported author can shape judgments of authenticity. We do not, however, have evidence that Hofstadter initially attributed a mind to the system and subsequently withdrew that attribution. His account documents a negative judgment about impersonation, which can be examined alongside the changes in attribution described in his philosophical essay.</p><p>A further clarification concerns the chronology of the two texts. The manuscript of <em>Is there an ‘I’ in AI?</em> is dated April 2023 and therefore precedes the article published in <em>The Atlantic</em> in July of the same year. Both reflections belong to a period in which Hofstadter was questioning the cognitive possibilities of the new models while also expressing concern about their ability to generate inauthentic accounts in a human voice.</p><p>The coexistence of these attitudes makes his case particularly interesting. Wonder at behavior that appears to reveal understanding exists alongside distrust of an artificial account claiming to express a personal history and voice. Recognizing possible cognitive abilities and judging the authenticity of a particular expression are distinct questions, whose relationship becomes especially clear in Hofstadter’s experience.</p><h3>The Sign, the Mind, and the Question That Remains Open</h3><p>In the concluding pages of <em>Is there an ‘I’ in AI?</em>, Hofstadter returns to a theme running through his entire body of work: the possibility that an “I” might emerge from complex structures of self-reference, supported by a physical substrate capable of representation and abstraction.</p><p>The notion of a <em>strange loop</em>, introduced in <em>Gödel, Escher, Bach</em> and developed further in <em>I Am a Strange Loop</em> (2007), provides the theoretical background for this perspective. Hofstadter conceives of the self as a structure through which a system represents its own characteristics and develops a model of itself. Within his philosophical framework, consciousness is closely connected to this self-referential organization.</p><p>From this perspective arises the possibility that future computational systems might develop sufficiently sophisticated forms of self-representation to make the recognition of a genuine artificial “I” plausible. Hofstadter writes:</p><blockquote><em>Computing systems with self-proclaimed ‘I’s’ will have real ‘I’s’ to the extent that they feel, to us, like our own ‘I’s’. That is to say, if and when a future AI’s use of the pronoun ‘I’ rings true to our ears, then it will be the sign of a real ‘I’ existing inside the machine.</em></blockquote><p>(Hofstadter, 2026, p. 9).</p><p>This passage advances a substantial philosophical claim. Hofstadter considers the possibility that coherent, stable, and self-referential linguistic behavior may provide grounds for recognizing the actual existence of an artificial self. His position rests on a conception of consciousness as an organized pattern, without granting an absolute ontological privilege to the biological substrate.</p><p>Noosemia offers a complementary perspective, focusing on how such an attribution becomes possible in the observer’s experience. When a system speaks about itself appropriately and maintains recognizable continuity in its behavior, an interlocutor may come to experience that behavior as a sign of subjectivity. Whether this experience provides sufficient evidence for the existence of artificial consciousness requires further philosophical arguments and independent empirical criteria.</p><p>Hofstadter’s accounts bring the distinction between experience and conclusion into particularly sharp focus. A seemingly appropriate answer may evoke the recognition of understanding, while a subsequent error may weaken it. The observer’s knowledge and expectations influence the judgment, as illustrated by the double standard Hofstadter himself acknowledges. During an interaction, the meaning assigned to artificial behavior may therefore change as the observer’s interpretation evolves.</p><p>The emotional dimension also plays a role in this process. In the stacking experiment, Hofstadter first describes his astonishment at a performance that seems to reveal understanding, followed by relief when an error reopens the question. In the autobiographical impersonation episode, his response is dominated by a rejection of inauthenticity. These reactions illustrate how encounters with artificial systems can acquire different emotional qualities, closely connected to the expectations brought to the interaction.</p><p>The scientific interest of these accounts does not depend on treating them as experimental proof of noosemia. They are qualitative testimonies, particularly valuable because they come from a scholar who has devoted much of his intellectual life to understanding cognitive processes. They make it possible to identify an interpretive pattern that may be investigated further through empirical research and longitudinal observation, with methods capable of distinguishing the felt strength of an attribution from its epistemic justification.</p><p>The speed of technological development makes this distinction all the more important. The models Hofstadter discusses belong to a particular stage in the evolution of generative artificial intelligence. His reflections retain their historical and philosophical value, while assessments of the systems’ capabilities must remain tied to the circumstances in which they were made. The possibility that later architectures may overcome previously observed limitations calls for judgments that remain open to revision, rather than turning the temporary capabilities of a model into definitive statements about what artificial intelligence can become.</p><p>There is something particularly significant about Hofstadter himself, who reflected on the dangers of the ELIZA effect decades ago, offering such a detailed account of his encounters with a new generation of language models. His experience reveals a complex relationship between technical knowledge and the recognition of cognitive abilities, in which wonder coexists with doubt and judgments are reconsidered in light of new observations.</p><p>Noosemia gives a name to the moment when something expressed or produced by a system comes to be recognized as a possible sign of mental activity. A-noosemia describes the subsequent weakening or withdrawal of that recognition when what is observed no longer supports the earlier inference. Together, these concepts help illuminate the dynamic character of mind attribution in human interactions with artificial intelligence.</p><p>In our relationships with other people, we routinely recognize minds through their words and actions. Contemporary AI introduces new, artificial interlocutors capable of producing behavior in domains long associated with human cognition, inviting us to examine more carefully the relationship between what we observe and what we infer.</p><p>Hofstadter’s experience thus acquires significance beyond the individual technological episodes. It invites us to consider how something we observe comes to be recognized as a sign of thought, and how that recognition may later be reconsidered. The question also concerns the way we make sense of other people, since attributing a mind is fundamental to how we experience and interpret the social world.</p><p><strong>Noosemia invites us to examine precisely this transition: the moment when, in the experience of an observer, a sign begins to appear as a sign of mind.</strong> Understanding how this happens, and what determines the strength or reversibility of that recognition, is a question that accompanies the progress of artificial intelligence while also reaching into our own experience as thinking beings.</p><h3>References</h3><p>Collins, E., &amp; Ghahramani, Z. (2021, May 18). <em>LaMDA: Our breakthrough conversation technology</em>. Google. <a href="https://blog.google/innovation-and-ai/products/lamda/">https://blog.google/innovation-and-ai/products/lamda/</a></p><p>De Santis, E. (2026, August 22). <em>Noosemia: L’esperienza del segno come segno di mente</em> [Noosemia: The experience of a sign as a sign of mind]. AION Research. <a href="https://www.aionresearch.it/2026/08/noosemia-lesperienza-del-segno-come.html">https://www.aionresearch.it/2026/08/noosemia-lesperienza-del-segno-come.html</a></p><p>De Santis, E., &amp; Rizzi, A. (2025). <em>Noosemia: Toward a cognitive and phenomenological account of intentionality attribution in human–generative AI interaction</em> [Preprint]. arXiv. <a href="https://doi.org/10.48550/arXiv.2508.02622">https://doi.org/10.48550/arXiv.2508.02622</a></p><p>Hofstadter, D. R. (1979). <em>Gödel, Escher, Bach: An eternal golden braid</em>. Basic Books.</p><p>Hofstadter, D. R. (1995). The ineradicable Eliza effect and its dangers. In <em>Fluid concepts and creative analogies: Computer models of the fundamental mechanisms of thought</em> (pp. 155–168). Basic Books.</p><p>Hofstadter, D. R. (2007). <em>I am a strange loop</em>. Basic Books.</p><p>Hofstadter, D. R. (2023, April). <em>Is there an “I” in AI?</em> [Manuscript]. <a href="https://berryvilleiml.com/wp-content/uploads/Is-there-an-%E2%80%9CI%E2%80%9D-in-AI-.pdf">https://berryvilleiml.com/wp-content/uploads/Is-there-an-%E2%80%9CI%E2%80%9D-in-AI-.pdf</a></p><p>Hofstadter, D. R. (2023, July 8). Gödel, Escher, Bach, and AI. <em>The Atlantic</em>. <a href="https://www.theatlantic.com/ideas/archive/2023/07/godel-escher-bach-geb-ai/674589/">https://www.theatlantic.com/ideas/archive/2023/07/godel-escher-bach-geb-ai/674589/</a></p><p>Hofstadter, D. R. (2026). Is there an ‘I’ in AI? <em>Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 384</em>(2320), Article 20240527. <a href="https://doi.org/10.1098/rsta.2024.0527">https://doi.org/10.1098/rsta.2024.0527</a></p><p>OpenAI. (2020, May 28). <em>Language models are few-shot learners</em>. <a href="https://openai.com/index/language-models-are-few-shot-learners/">https://openai.com/index/language-models-are-few-shot-learners/</a></p><p>OpenAI. (2023, March 14). <em>GPT-4</em>. <a href="https://openai.com/index/gpt-4-research/">https://openai.com/index/gpt-4-research/</a></p><p>OpenAI. (2025, April 10). <em>ChatGPT release notes</em> [Announcement of GPT-4’s retirement from ChatGPT, effective April 30, 2025]. <a href="https://help.openai.com/en/articles/6825453-chatgpt-release-notes">https://help.openai.com/en/articles/6825453-chatgpt-release-notes</a></p><p>OpenAI. (2026). <em>Deprecations</em> [API documentation, accessed October 16, 2026]. <a href="https://developers.openai.com/api/docs/deprecations">https://developers.openai.com/api/docs/deprecations</a></p><p>Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., &amp; Polosukhin, I. (2017). Attention is all you need. <em>Advances in Neural Information Processing Systems, 30</em>.</p><p>Weizenbaum, J. (1966). ELIZA — A computer program for the study of natural language communication between man and machine. <em>Communications of the ACM, 9</em>(1), 36–45. <a href="https://doi.org/10.1145/365153.365168">https://doi.org/10.1145/365153.365168</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0e95d8edb4df" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond Next-Token Prediction. How Transformers architectures can learn rules]]></title>
            <link>https://medium.com/@enrico.desantis/beyond-next-token-prediction-848201d46893?source=rss-e9539446c6d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/848201d46893</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[representation-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[transformers]]></category>
            <category><![CDATA[mathematics]]></category>
            <dc:creator><![CDATA[Enrico De Santis]]></dc:creator>
            <pubDate>Sat, 12 Sep 2026 08:18:58 GMT</pubDate>
            <atom:updated>2026-09-12T09:22:28.159Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AiJe2cP-ruuTN2mJDwSwew.png" /></figure><p><em>Transformers can learn internal rules and procedures, going beyond simple memorization and next-token prediction.</em></p><p>Saying that a Large Language Model learns to predict the next token is correct. Yet this description remains profoundly insufficient to explain what the model has actually learned and what computational structures it relies on when producing an answer.</p><p>A preliminary clarification is useful. Today, an LLM is often only one component within a much broader computational infrastructure, which may include agentic orchestration, external memory, retrieval systems, analytical tools, executable code, specialized skills, and access to external data sources. These capabilities deserve a separate analysis. Here, however, I deliberately place them in the background and focus on a more specific question: what can emerge inside the neural model itself, even when no external tool is available at inference time?</p><p>Without this broader perspective, it becomes difficult to understand how an LLM can perform mathematical computations and, more generally, exhibit forms of structured algorithmic reasoning.</p><p>Popular accounts often suggest that, during training, a neural model simply «memorizes the data» and then, like a «stochastic parrot», regurgitates them at inference time. The expression «stochastic parrot» has a specific history and meaning in the critical literature on Large Language Models (Bender et al., 2021). As an aside, actual parrots are vastly more intelligent than we commonly assume. In any case, a purely mnemonic account cannot provide a general explanation of how neural networks operate.</p><p>Similarly, describing a neural model as a «statistical engine» is sometimes taken to mean that its answers ultimately amount to something akin to random coin tossing. Statistics is certainly there, but stable representations and structured computational procedures can also emerge within the network (Nanda et al., 2023; Maltoni &amp; Ferrara, 2024).</p><p>During autoregressive pretraining, the training objective specifies the task, which we may simplify as «predict the next token». The Transformer architecture and the dynamics of learning determine what internal representations and procedures can emerge in order to accomplish that task.</p><p>A particularly instructive example comes from the scientific literature on «grokking» (Power et al., 2022; Nanda et al., 2023).</p><p>For simplicity, consider a small Transformer trained on a basic mathematical operation such as addition modulo 10. In this arithmetic, 7 + 8 = 5 because, once 10 is reached, counting starts again from 0, as shown in the figure. Classical grokking experiments use analogous problems in modular arithmetic.</p><p>The key point is that, during training, the Transformer sees only many examples consisting of input-output pairs, within an entirely data-driven learning paradigm. It may observe, for instance, 2 + 3 → 5 or 7 + 8 → 5. No one provides it with the general rule of modular addition, nor with an algorithm explaining how the result should be obtained. What we may informally call the «meta-rule» must emerge from the structure of the examples themselves.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F5eU92IaSqOcloKvUA_0dg.png" /></figure><p>This is where things become interesting. During an initial phase, the model may learn the observed pairs perfectly while continuing to fail on new ones. Its behavior is then largely compatible with memorization of the training data. As training continues, however, a striking change may occur. The network develops an internal structure capable of capturing the mathematical regularity of the problem and begins to generalize correctly to input-output combinations it has never encountered during training (Power et al., 2022).</p><p>This is the phenomenon known as «grokking».</p><p>Viewed from the outside, generalization may appear relatively suddenly after a prolonged period of overfitting. Mechanistic interpretability has allowed researchers to look inside this transition. Nanda et al. (2023) reconstructed the mechanism learned by small Transformers performing modular addition, revealing the progressive formation of computational circuits based on Fourier representations and trigonometric identities. They distinguish a memorization phase, a circuit-formation phase, and a subsequent «cleanup» phase in which components associated with memorization are progressively removed.</p><p>The network can therefore arrive at the result of an operation by executing a computational procedure that emerged during training, without requiring that the particular input-output pair had previously been memorized. Other experiments on arithmetic have further shown that language models trained through next-token prediction can generalize to arithmetic instances and input combinations not encountered during training, while developing internal representations in which numerical computation takes place (Maltoni &amp; Ferrara, 2024).</p><p>Understanding these systems therefore requires us to look inside the model as well, examining what representations are formed, what regularities are captured, and what procedures are applied during inference. Next-token prediction describes the training objective. Explaining the resulting behavior requires another level of analysis.</p><p>Today, in controlled settings, neural networks and Transformers have been shown to learn abstract patterns and algorithmic procedures that can be reapplied to novel inputs, including forms of generalization beyond the sequence lengths encountered during training (Zhou et al., 2024). Looking ahead, as model capacity increases and models become increasingly able to compose previously learned procedures, this points toward the possibility of acquiring, combining, and applying an extraordinarily large variety of algorithms.</p><p>This perspective also helps explain why highly advanced neural systems can be employed in mathematical research on problems of extraordinary complexity. A very recent example is the solution to the Navier–Stokes existence and smoothness problem announced by OpenAI on September 8, 2026, reportedly produced by an internal system and accompanied by a formalization in Lean (OpenAI, 2026). This is a claim of exceptional significance that is currently being scrutinized by the mathematical community and should therefore be regarded with the appropriate caution (Castelvecchi, 2026).</p><p>P.S. If a network genuinely learns a correct rule and applies it correctly, the validity of the result does not depend on whether that particular answer appeared in the training data. The epistemic question then shifts from the mnemonic provenance of the answer to the correctness of the learned rule and of its application.</p><p>References</p><p>Bender, E. M., Gebru, T., McMillan-Major, A., &amp; Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? 🦜. In <em>Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency</em> (pp. 610–623). Association for Computing Machinery. <a href="https://doi.org/10.1145/3442188.3445922">https://doi.org/10.1145/3442188.3445922</a></p><p>Castelvecchi, D. (2026, September 8). OpenAI claims huge maths breakthrough on a famed ‘Millennium Problem’. <em>Nature</em>. <a href="https://doi.org/10.1038/d41586-026-02842-5">https://doi.org/10.1038/d41586-026-02842-5</a></p><p>Maltoni, D., &amp; Ferrara, M. (2024). Arithmetic with language models: From memorization to computation. <em>Neural Networks, 179</em>, 106550. <a href="https://doi.org/10.1016/j.neunet.2024.106550">https://doi.org/10.1016/j.neunet.2024.106550</a></p><p>Nanda, N., Chan, L., Lieberum, T., Smith, J., &amp; Steinhardt, J. (2023). Progress measures for grokking via mechanistic interpretability. In <em>The Eleventh International Conference on Learning Representations</em>. <a href="https://openreview.net/forum?id=9XFSbDPmdW">https://openreview.net/forum?id=9XFSbDPmdW</a></p><p>OpenAI. (2026, September 8). <em>On the Navier–Stokes Millennium Prize Problem</em>. OpenAI. <a href="https://openai.com/index/navier-stokes-solution/">https://openai.com/index/navier-stokes-solution/</a></p><p>Power, A., Burda, Y., Edwards, H., Babuschkin, I., &amp; Misra, V. (2022). <em>Grokking: Generalization beyond overfitting on small algorithmic datasets</em>. arXiv. <a href="https://doi.org/10.48550/arXiv.2201.02177">https://doi.org/10.48550/arXiv.2201.02177</a></p><p>Zhou, H., Bradley, A., Littwin, E., Razin, N., Saremi, O., Susskind, J. M., Bengio, S., &amp; Nakkiran, P. (2024). What algorithms can Transformers learn? A study in length generalization. In <em>The Twelfth International Conference on Learning Representations</em>. <a href="https://openreview.net/forum?id=RtS4JQ4K0P">https://openreview.net/forum?id=RtS4JQ4K0P</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=848201d46893" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Real Leap of Transformers Is Learning How to Represent Context]]></title>
            <link>https://medium.com/@enrico.desantis/the-real-leap-of-transformers-is-learning-how-to-represent-context-5a1700bf03d2?source=rss-e9539446c6d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/5a1700bf03d2</guid>
            <category><![CDATA[representation-learning]]></category>
            <category><![CDATA[transformers]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Enrico De Santis]]></dc:creator>
            <pubDate>Sat, 12 Sep 2026 07:40:07 GMT</pubDate>
            <atom:updated>2026-09-12T08:06:08.253Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GfcjtgaDuPVUxB4rU0ba1A.png" /></figure><p><em>Why next-token prediction became dramatically more powerful when language models learned the space in which probability is estimated</em></p><p>Large Language Models are often described, correctly, as systems trained to predict the next token. Taken on its own, however, this description conceals much of what makes contemporary language models remarkable.</p><p>Next-token prediction is far from a recent invention. Language modeling has long been formulated as a probabilistic problem. Given a sequence of tokens x₁, x₂, …, xₜ, its joint probability can be factorized as</p><p>p(x₁, …, xₜ) = ∏ᵀₜ₌₁ p(xₜ ∣ x₁, …, xₜ₋₁)</p><p>An autoregressive model therefore learns to estimate the probability of each token conditioned on what precedes it. Modern language models still solve essentially this problem. During training, their parameters are adjusted so that the probability assigned to the observed continuation increases, usually through minimization of cross-entropy or, equivalently, negative log-likelihood.</p><p>The decisive change introduced by neural language modeling concerned the representation of the information on which that probability is conditioned. This distinction offers a useful starting point for understanding why modern language models are so much more powerful than earlier statistical approaches.</p><h3>Probability Was There from the Beginning</h3><p>Classical statistical language models estimate quantities such as</p><p>p(xₜ ∣ xₜ₋ₙ₊₁, …, xₜ₋₁)</p><p>An n-gram model represents context through a fixed number of preceding symbols. A trigram model, for example, predicts the next word using the two immediately preceding words.</p><p>The underlying idea is elegant and effective. If a sequence has been observed many times, its empirical frequency provides evidence about likely continuations. Decades of research produced sophisticated methods for handling sparse observations, including back-off, interpolation and several forms of smoothing. Katz’s back-off model and the systematic analyses later carried out by Chen and Goodman remain landmarks in this tradition (Katz, 1987; Chen &amp; Goodman, 1999).</p><p>Yet the fundamental difficulty is combinatorial. If the vocabulary contains V symbols, the number of possible sequences of length n grows as Vⁿ. Even for moderate values of n, the overwhelming majority of possible contexts will never occur in the training corpus. As the conditioning context grows, exact observations rapidly become rare.</p><p>The difficulty therefore concerns more than the amount of available data. It also concerns how linguistic experience is represented. Two sentences may express closely related structures while sharing very few surface-level sequences. A model based primarily on discrete configurations has limited opportunities to transfer evidence from one to the other unless additional assumptions or smoothing mechanisms are introduced.</p><p>This is the point at which learned representations changed the nature of the problem.</p><h3>From Discrete Configurations to Learned Spaces</h3><p>Bengio, Ducharme, Vincent and Jauvin made this transition particularly clear in their 2003 neural probabilistic language model. Their proposal addressed the curse of dimensionality by learning distributed representations of words jointly with the probability function used for prediction (Bengio et al., 2003).</p><p>Instead of representing each word solely as an unrelated symbolic identity, the model maps words into a continuous space. Schematically,</p><p>word → distributed representation → probability distribution</p><p>Words that behave similarly can acquire related representations, while the parameters used to estimate probabilities are shared across many linguistic situations. The resulting mechanism of generalization differs fundamentally from one based only on discrete counts.</p><p>A sequence that has never appeared during training need not be treated as an entirely new event. Its components and their relationships may generate a representation related to those associated with previously observed contexts. Statistical evidence can therefore be exploited across contexts that differ at the symbolic level while sharing useful structure in representation space.</p><p>The principle can be written schematically as</p><p>hₜ = fθ(x₁, …, xₜ₋₁)</p><p>followed by</p><p>pθ(xₜ ∣ x₁, …, xₜ₋₁) = softmax(Whₜ + b)</p><p>The quantity hₜ is the essential bridge between context and probability. It represents the preceding sequence in a form learned from data.</p><p>The learning problem has consequently changed in an important way. The model must still estimate which token is likely to follow, while also learning which properties of the preceding sequence should matter for that estimation. Representation has become part of the probability-estimation problem itself.</p><h3>Before Transformers</h3><p>This transition began well before Transformers.</p><p>Recurrent neural networks extended neural language modeling by constructing representations that could incorporate information from increasingly long preceding sequences. Mikolov and colleagues showed that recurrent language models could substantially reduce perplexity relative to strong back-off baselines, providing early empirical evidence for the value of continuous representations and learned sequential context (Mikolov et al., 2010).</p><p>By this stage, the central idea was already established. A language model could generalize more effectively when it learned a distributed description of the past rather than relying exclusively on discrete counts over fixed contexts.</p><p>Transformers expanded this principle by changing the way such contextual representations are constructed.</p><h3>What Attention Changes</h3><p>The Transformer introduced a mechanism in which contextual representations can be built through direct, content-dependent interactions among positions in a sequence (Vaswani et al., 2017).</p><p>Through self-attention, the representation associated with one position can depend on representations associated with many other positions. In simplified form,</p><p>hᵢ = Σⱼ αᵢⱼvⱼ</p><p>where the coefficients αᵢⱼ depend on learned relations between the current query and the available keys.</p><p>The relevance of a previous token therefore varies with context. A word several positions away may be highly informative in one sentence and almost irrelevant in another. Different attention heads can learn different interaction patterns, while successive layers repeatedly transform the resulting representations.</p><p>The conditioning information is thus reconstructed as the sequence moves through the network. From this perspective, a Transformer does more than estimate</p><p>p(next token ∣ previous tokens)</p><p>It learns a sequence of transformations that determines how those previous tokens should be represented before the probability distribution is produced.</p><p>That distinction is subtle, but it changes the expressive capacity of the model substantially.</p><h3>When Context Becomes Geometry</h3><p>Once context is represented in a learned continuous space, relationships that are difficult to express through discrete counts can emerge geometrically.</p><p>Two different sequences may produce internal states that share relevant structure. Their surface forms may differ considerably, while their representations reflect similar syntactic dependencies, semantic roles or contextual relationships.</p><p>Research on contextual representations provides substantial evidence that these internal spaces acquire structured organization.</p><p>Hewitt and Manning showed that syntactic tree relationships can be recovered from the geometry of contextual word representations through a relatively simple learned transformation (Hewitt &amp; Manning, 2019). Their result does not imply that a Transformer literally stores symbolic parse trees. It shows that syntactic structure is systematically reflected in the geometry of the learned space.</p><p>Ethayarajh examined BERT, ELMo and GPT-2 and showed how strongly contextualized their representations become across layers. The same lexical item can occupy substantially different regions of representation space depending on the context in which it occurs (Ethayarajh, 2019).</p><p>This observation is particularly significant for language modeling. The object from which the next-token distribution is computed is no longer merely the identity of the preceding words. It is a contextual state constructed from them.</p><p>The conditional distribution is therefore evaluated through a representation whose geometry has itself been learned.</p><h3>Representation Changes Generalization</h3><p>This shift also changes the way we should think about generalization.</p><p>Consider two sentences that have never appeared verbatim in the training data. Their wording may differ, their lengths may differ, and the relevant entities may be entirely different. Yet both may instantiate the same grammatical relation, semantic dependency or recurring computational pattern.</p><p>An n-gram model has limited direct mechanisms for recognizing such equivalence because its probability estimates are closely tied to discrete local configurations. A neural model can, in principle, produce internal representations that share some relevant structure.</p><p>Generalization can then exploit distances, directions and transformations within the learned space.</p><p>This is one reason why describing an LLM simply as a system that predicts the next token can be misleading. The statement identifies the training objective while remaining largely silent about the machinery learned to accomplish it.</p><p>The distinction is similar to saying that a classifier predicts labels. This tells us what the system outputs, yet tells us much less about the features, latent variables or decision geometry constructed during learning.</p><p>With contemporary language models, the difference becomes particularly important because the internal computation is deep, distributed and repeatedly contextualized.</p><h3>Prediction Can Induce Internal Structure</h3><p>The training objective does not explicitly instruct the model to learn syntax, semantic relationships or particular computational mechanisms. It rewards accurate prediction.</p><p>Across a sufficiently rich distribution of linguistic data, however, internal structure can become useful precisely because it improves prediction.</p><p>A model that must predict language in many different contexts benefits from representing distinctions that systematically influence what comes next. Grammatical dependencies constrain possible continuations. Semantic relations alter expectations about subsequent words. Entity identity matters because references recur. Structured forms of computation may likewise become useful when they improve prediction across classes of related examples.</p><p>Research into Transformer internals provides evidence compatible with this picture.</p><p>Geva and colleagues showed that Transformer feed-forward layers can be interpreted as key-value memories in which learned patterns progressively interact with the model’s prediction process. They also found a tendency for lower layers to capture shallower patterns while higher layers exhibit more semantic structure (Geva et al., 2021).</p><p>Such results should be interpreted with appropriate caution. They do not amount to a complete theory of Transformer computation. They nevertheless support an important broader conclusion. A comparatively simple predictive objective can give rise to rich internal representations when those representations improve prediction across many contexts.</p><h3>The Objective and the Learned Computation</h3><p>This distinction is worth making explicit.</p><p>The training objective specifies what counts as success.</p><p>The learned computation determines how that success is achieved.</p><p>For an autoregressive language model, the objective can remain mathematically simple even when the internal mechanism required to optimize it becomes highly structured.</p><p>This is why the expression “next-token predictor” is simultaneously accurate and incomplete. It correctly identifies the probabilistic objective, while leaving the representational and computational structures that emerge during learning largely unspecified.</p><p>Confusing these two levels easily produces an impoverished picture of contemporary language models.</p><h3>The Deeper Transition</h3><p>The history of language modeling can therefore be read as a progressive transformation of the conditioning mechanism.</p><p>Classical n-gram models estimate probabilities over explicitly defined discrete contexts. Neural language models learn continuous representations of those contexts. Recurrent models allow such representations to evolve sequentially. Transformers repeatedly reconstruct contextual information through content-dependent interactions across the sequence.</p><p>Probability remains central throughout this history. What changes is the space in which the relevant evidence is represented and organized.</p><p>This provides a more precise description of the transition from traditional statistical language modeling to modern Transformers. The decisive advance was the realization that the representation on which statistical prediction depends could itself be learned.</p><p>Once representation becomes part of the learning problem, the model can discover similarities, dependencies and abstractions that were not explicitly encoded in the original symbolic description of the data. Statistical learning supplies the signal through which this organization emerges, while the resulting representation determines how broadly the acquired regularities can be reused across contexts.</p><p>The remarkable capabilities of modern language models arise, to a significant extent, from this interaction between probability and representation.</p><h3>References</h3><p>Bengio, Y., Ducharme, R., Vincent, P., &amp; Jauvin, C. (2003). A neural probabilistic language model. <em>Journal of Machine Learning Research, 3</em>, 1137–1155.</p><p>Chen, S. F., &amp; Goodman, J. (1999). An empirical study of smoothing techniques for language modeling. <em>Computer Speech &amp; Language, 13</em>(4), 359–394. <a href="https://doi.org/10.1006/csla.1999.0128">https://doi.org/10.1006/csla.1999.0128</a></p><p>Ethayarajh, K. (2019). How contextual are contextualized word representations? Comparing the geometry of BERT, ELMo, and GPT-2 embeddings. <em>Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing</em>, 55–65. <a href="https://doi.org/10.18653/v1/D19-1006">https://doi.org/10.18653/v1/D19-1006</a></p><p>Geva, M., Schuster, R., Berant, J., &amp; Levy, O. (2021). Transformer feed-forward layers are key-value memories. <em>Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing</em>, 5484–5495. <a href="https://doi.org/10.18653/v1/2021.emnlp-main.446">https://doi.org/10.18653/v1/2021.emnlp-main.446</a></p><p>Hewitt, J., &amp; Manning, C. D. (2019). A structural probe for finding syntax in word representations. <em>Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1</em>, 4129–4138. <a href="https://doi.org/10.18653/v1/N19-1419">https://doi.org/10.18653/v1/N19-1419</a></p><p>Katz, S. M. (1987). Estimation of probabilities from sparse data for the language model component of a speech recognizer. <em>IEEE Transactions on Acoustics, Speech, and Signal Processing, 35</em>(3), 400–401. <a href="https://doi.org/10.1109/TASSP.1987.1165125">https://doi.org/10.1109/TASSP.1987.1165125</a></p><p>Mikolov, T., Karafiát, M., Burget, L., Černocký, J., &amp; Khudanpur, S. (2010). Recurrent neural network based language model. <em>Proceedings of Interspeech 2010</em>, 1045–1048. <a href="https://doi.org/10.21437/Interspeech.2010-343">https://doi.org/10.21437/Interspeech.2010-343</a></p><p>Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., &amp; Polosukhin, I. (2017). Attention is all you need. <em>Advances in Neural Information Processing Systems, 30</em>, 5998–6008.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5a1700bf03d2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Noosemia and Symbolic Experience: When the Sign Seems to Have a Mind]]></title>
            <link>https://medium.com/@enrico.desantis/noosemia-and-symbolic-experience-when-the-sign-seems-to-have-a-mind-37a4c709d95c?source=rss-e9539446c6d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/37a4c709d95c</guid>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[philosophy-of-mind]]></category>
            <category><![CDATA[noosemia]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Enrico De Santis]]></dc:creator>
            <pubDate>Mon, 07 Sep 2026 17:35:22 GMT</pubDate>
            <atom:updated>2026-09-07T17:35:22.783Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Black-and-white ink illustration of a human and a humanoid robot joining two complementary halves of a circular symbol. The image evokes noosemia by showing meaning emerging not only from the completed sign, but from the apparent relation between expression, source, and mind." src="https://cdn-images-1.medium.com/max/1024/1*vocIl-Hg5H-i57dXNV-BFw.png" /></figure><p><em>What happens when language no longer feels like mere output, but begins to feel as if it comes from someone? This essay explores </em><strong><em>noosemia</em></strong><em>: the experience of a sign as a sign of mind. Moving from Plato and Jung to Large Language Models, it asks how meaning can exceed content and begin to suggest a source, an interiority, a possible mind. The core proposal is simple: in human–AI interaction, an utterance may point not only to </em><strong><em>what is said</em></strong><em>, but also to </em><strong><em>who seems to be saying it</em></strong><em>. That second movement is where noosemia begins.</em></p><blockquote>“Another consciousness may also regard the mathematical expression as a symbol of an unknown psychic state of affairs concealed within the intention behind its production.”<br> — C. G. Jung, Psychological Types (1921), CW 6, para. 817; trans. authors from the German.</blockquote><p>On September 3, 2026, OpenAI introduced <strong>GPT-6 Astra</strong>, a model that marks a further advance in the capacity of generative systems to tackle complex tasks spanning language, research, programming, computer use, and interaction with digital environments (OpenAI, 2026). The interest surrounding its release naturally concerns performance, operational autonomy, the scope for control, and the growing difficulty of making systems of this complexity fully transparent. Alongside these questions, however, another issue is coming into view, one less readily captured by measurement and more directly bound up with the<strong> experience of interaction</strong>. When a system preserves the thread of a conversation, grasps an implication, revises a strategy, coordinates different tools, or redirects its behaviour in response to what happens during a task, the system’s individual performances tend to coalesce, in the eyes of the interlocutor, into a unified configuration. The relevance of a response may then take on the <em>form of</em> understanding, continuity of context that of memory, strategic revision that of reconsideration, and the choice among alternatives that of judgement. At this point the nature of the question changes:<em> a sequence of observable behaviours begins to suggest a source from which they seem to arise, and through language the indirectly present and still indeterminate figure of a possible mental source begins to take shape</em>.</p><p>The distance between these two modes of experience is minimal linguistically and vast phenomenologically. In the first case, we describe observable properties of the utterance; in the second, the profile of a possible speaker begins to emerge through the utterance itself. The text continues to mean something and, at the same time, <strong>begins to appear plausibly</strong> as the expression of someone; what seems to express itself through it gradually takes on the features of a source capable of understanding, remembering, evaluating, responding, and perhaps adopting a perspective.</p><p>It is in the gap between the observable properties of the utterance and the phenomenological appearance of a <strong><em>possible</em> mental source</strong> that what we have called noosemia takes shape. The term designates a cognitive and phenomenological mode of experience in which, during interaction with generative systems — especially dialogic or multimodal ones — the human interlocutor tends to attribute intentionality, agency, mental states, or some form of interiority to the system. Such<em> attribution</em> describes the interlocutor’s experience while leaving open the question of whether artificial consciousness is actually present. <a href="https://medium.com/@enrico.desantis/noosemia-perceiving-the-mind-in-signs-7b23a14c830e"><strong>Noosemia</strong></a> may emerge independently of bodily resemblance to a human being, particularly where <em>linguistic performance</em>, <em>epistemic opacity</em>, and the <em>systemic complexity of the interaction</em> make the intuitive boundary between formally competent behaviour and the manifestation of a mental source increasingly unstable (De Santis &amp; Rizzi, 2025).</p><p>The decisive question arises at a level logically prior to the truth of the attribution. Before asking whether a <strong>Large Language Model</strong> (LLM) <em>genuinely</em> understands, whether it possesses <em>representations</em> comparable to mental states, or whether notions such as <em>intentionality</em> and <em>subjectivity</em> can be applied to artificial systems, we must ask <strong>how a configuration of signs can be experienced as the manifestation of a mind</strong>, even when the machine makes a mistake. What transformation must the sign undergo for it to preserve its signifying function while also beginning to point towards someone?</p><p>To understand this passage, we must temporarily leave artificial intelligence behind and follow a much older history, in which the “sign” began to exceed what it immediately signifies.</p><h4>1. The Part and What Is Missing</h4><p>Before it came to designate more generally an emblem, an image, or a conventional sign, <em>sýmbolon</em> referred to a relation. Greek lexicography records σύμβολον as, among other things, either of two matching halves or corresponding parts of an object broken between guests or contracting parties and preserved so that their later reunion could enable mutual recognition (Liddell et al., 1940, s.v. σύμβολον). Neither half constituted the whole on its own; precisely through its incompleteness, each preserved a reference to the other. Recognition occurred when the two parts were brought together again and the fracture revealed itself, retrospectively, as the trace of a belonging.</p><p>In the Symposium, 191d, Plato significantly uses the same word when Aristophanes says that each of us is a xýmbolon of a human being, a separated part in search of its counterpart: “Ἕκαστος οὖν ἡμῶν ἐστὶν ἀνθρώπου ξύμβολον” (Hékastos oûn hēmôn estin anthrṓpou xýmbolon; “Each of us, then, is a xýmbolon of a human being”, our translation) (Plato, Symposium, 191d). The later history of the word extends far beyond this image, yet a structure destined for a long theoretical afterlife is already visible within it. The symbol first appears as an incomplete presence which, through its very lack of <strong>self-sufficiency</strong>, points towards what is missing and preserves the possibility of recomposition.</p><p>The philosophical reading of <em>sym-bállein</em>, understood as bringing together, joining, or gathering, has emphasised this capacity to hold together what separation has placed at a distance. Among contemporary interpretations, Galimberti develops the polarity between sym-bállein and dia-bállein, between composition and separation, within a genealogy that moves from a symbolic universe capable of accommodating ambivalence towards a conceptual order in which each thing tends to be determined according to identity and non-contradiction (Galimberti, 1984/2001, pp. 58–59).</p><p>The relation between symbol and the <strong>principle of non-contradiction</strong> requires some clarification. In Aristotle’s formulation, the same attribute cannot belong and fail to belong to the same thing at the same time and in the same respect (Aristotle, Metaphysics, Γ 3, 1005b19–20). <em>Symbolic ambivalence</em> operates at the <strong>level of signification</strong> and leaves this logical constraint intact: a determination may preserve its identity while entering into a further constellation of relations. What is A remains A and may also point towards B.</p><p>The ritual example of the Wachandi, taken up in reflections on symbolism within the Jungian literature, makes this distinction visible. The <strong><em>pit</em></strong> retains its material identity as a pit while, within the ritual, also taking on the value of a <strong><em>vulva</em></strong>; the first determination thereby ceases to exhaust the totality of the relations of sense in which it may participate (Galimberti, 1984/2001, pp. 58–59). If the regime of the sign tends towards A → A, the symbol preserves the possibility A ⇝ {A, B, …}. Symbolic excess consists in opening sense to further determinations compatible with the <strong>identity of the referent</strong>.</p><p>A theoretical structure thus begins to appear that extends beyond the particular genealogy of the West within which this interpretation is situated. The symbol entails a relation between presence and absence, between what is given and what participates in the configuration of sense while withdrawing from the same mode of presence. Its elementary movement may be expressed as: <strong>presence → lack → reference → possible recomposition</strong>.</p><p>In the ancient sýmbolon, <strong><em>absence</em></strong> could still be imagined as the other half of the token. With Jung, many centuries later, the structure becomes more radical: what the symbol points towards may be something only partially known, still awaiting an adequate formulation (Jung, 1921/1971, CW 6, paras. 817–819).</p><p>In Jung’s <em>Psychological Types</em> (1921/1971, CW 6, paras. 817–819), the distinction between sign and symbol lies primarily in the relation between what is already known and what, although manifested in experience, remains only <strong>partially known</strong>. The sign points to something that can be sufficiently determined; the symbol preserves an excess beyond the determination currently available.</p><p>Jung formulates this principle particularly clearly in § 897 (CW 6, para. 817):</p><p><em>An expression put forward for a known thing always remains a mere sign and is never a symbol. It is therefore quite impossible to create a living symbol — that is, one pregnant with meaning — out of known connections, for what is created in this way never contains more than what has been put into it. Every psychic product, insofar as it is at that moment the best possible expression for a state of affairs that is still unknown or only relatively known, can be understood as a symbol, provided one is disposed to assume that the expression is also intended to designate what is only divined but not yet clearly known. Insofar as every scientific theory contains a hypothesis and is therefore an anticipatory designation of a state of affairs that is still essentially unknown, it is a symbol. Furthermore, every psychological phenomenon is a symbol on the assumption that it says or signifies something more and something different that eludes present knowledge. This assumption is possible wherever there is a consciousness attuned to further possibilities of meaning in things. It is not possible only where — and only for that same consciousness — it has itself produced an expression intended to say exactly what the intention behind its production meant it to say, as, for example, in a mathematical expression. For another consciousness, however, this limitation does not apply at all. It may also understand the mathematical expression as a symbol for an unknown psychic state of affairs concealed within the intention that produced it, provided it can be shown that this state of affairs was unknown to the person who created the semiotic expression and therefore could not have been the object of conscious use.</em></p><p><em>(Jung, 1921; trans. authors.)</em></p><p>In this definition, the unknown is the very condition of the symbol’s vitality. As long as no more adequate formulation exists for what the expression points towards, the symbol remains the best available form of something that still awaits fuller <strong>determination</strong>. Its structure therefore takes the form s ⇝ X?: the question mark preserves the further dimension that present determination still leaves open.</p><p>The question then shifts decisively from the expression to the consciousness that encounters it (§ 898; CW 6, para. 818):</p><p><em>Whether something is a symbol or not depends first of all on the attitude of the observing consciousness: for example, on that of an intellect that regards the given fact not merely as such but also as an expression of something unknown. It is therefore entirely possible that someone may produce a fact that does not appear symbolic at all to them, yet does appear symbolic to another consciousness. The reverse is equally possible. There are, however, products whose symbolic character does not depend solely on the attitude of the observing consciousness, but manifests itself of its own accord in a symbolic effect upon the observer. Such products are constituted in such a way that they would have to be devoid of any meaning if they did not possess a symbolic meaning. A triangle with an eye enclosed within it is, as bare actuality, so meaningless that the observer cannot possibly regard it as mere accidental play. Such a configuration immediately compels a symbolic interpretation. This effect is reinforced either by the repeated and identical occurrence of the same configuration or by the particularly careful manner of its production, which is itself an expression of the special value placed upon it.</em></p><p><em>(Jung, 1921; trans. authors.)</em></p><p>This movement leads to a <strong>relational conception of symbolicity</strong>. The properties of the phenomenon and the disposition of consciousness jointly shape what occurs. The same phenomenon may acquire <strong>symbolic value</strong> for one consciousness and lose it for another, while certain configurations exert what Jung calls a “<strong><em>symbolic effect</em></strong>” on the <em>observer</em>.</p><p>Jung gives this disposition a name in § 899 (CW 6, para. 819):</p><p><em>Symbols that do not exert an effect of themselves in the way just described are either dead — that is, superseded by a better formulation — or products whose symbolic nature depends exclusively on the attitude of the observing consciousness. We may call this attitude, which apprehends the given phenomenon as symbolic, the symbolic attitude. It is justified only in part by the behaviour of things themselves; in another part it is the outgrowth of a particular worldview, one that attributes meaning to events, whether great or small, and places a certain greater value on this meaning than on bare actuality. Opposed to this view is another that always places the emphasis on bare actuality and subordinates meaning to facts. For this latter attitude there are no symbols wherever symbolism depends exclusively on the mode of observation. Yet there are symbols even for it: namely those that invite the observer to suppose the existence of a hidden meaning. An image of a bull-headed god can indeed be explained as a human body with a bull’s head on it. But this explanation can hardly hold its own against the symbolic explanation, for the symbol is too insistent to be overlooked.</em></p><p><em>(Jung, 1921; trans. authors.)</em></p><p>The symbolic attitude preserves factuality and keeps interpretation within constraints, yet suspends the claim that what has been determined in a phenomenon must coincide with the totality of what that phenomenon may mean (Jung, 1921/1971, CW 6, paras. 817–819). Its principle might be condensed into the formula present datum ≠ totality of possible sense.</p><p>Within this opening, the double meaning of “<strong>sense</strong>” becomes especially relevant. <strong>Sense designates both what something means and the direction in which one proceeds.</strong> The distinction between <em>Bedeutung</em> and <em>Sinn</em> articulates this difference (Galimberti, 1984/2001, p. 104): meaning is associated with intellectual determination, whereas sense also preserves the dimension of direction and reference. In the gap between the two lies the distance between <strong>rational</strong> and <strong>symbolic vision</strong>, the latter promoting “excess, furtherness, and reference” (Galimberti, 1984/2001, p. 104; trans. authors).</p><p>Within the regime of the sign, the prevailing form is s → m; within the symbolic regime, by contrast, we find s ⇝ m + Δm, where Δm indicates the excess that keeps meaning open to further participation in sense. A rigid decoding — such as one that transformed the bell tower in the dream, in a Freudian manner, into a conventional equivalent of the phallus (Galimberti, 1984/2001, p. 208) — would reduce the symbol to a fixed semiotic correspondence. In this reading of the Jungian distinction, symbolic reference proceeds “from present sense to a further participation in sense” (Galimberti, 1984/2001, p. 208; trans. authors).</p><p>The symbol preserves present meaning while opening it towards a totality of sense that exceeds it.</p><h4>2. The Excess of Signification</h4><p>If Jung encounters excess through the <strong>inexhaustibility of expression</strong>, anthropology encounters it through the disproportion between what is available for signification and what, at a given moment, has already been signified.</p><p>The case of mana, discussed by Henri Hubert and Marcel Mauss in their Esquisse d’une théorie générale de la magie, presents a polyvalent notion: it may qualify an object, the magician, the rite, and even the formula, and is described as an unfixed force distributed among beings, things, and events (Hubert &amp; Mauss, 1902–1903; Mauss, 1972). Of particular relevance is the resistance such a configuration offers to any perfectly univocal correspondence between signifier and signified.</p><p>Claude Lévi-Strauss, in Introduction to the Work of Marcel Mauss, turns this difficulty into a general theoretical problem. A disproportion persists between the available signifier and the mapped-out signified. Lévi-Strauss describes a “<strong>signifier-surfeit</strong>” which, in the human effort to understand the world, becomes a “<strong>surplus of signification</strong>”. The distribution of this “supplementary ration” allows the available signifier and the mapped-out signified to remain in the relation of complementarity that constitutes the very condition for the exercise of <strong>symbolic thought</strong> (Lévi-Strauss, 1950/1987, pp. 62–63).</p><p>The symbolic therefore inhabits the gap between the available signifier and the mapped-out signified. If every signifier possessed a single, fully determined signified and every signified had a perfectly equivalent formulation, the semiotic universe would function as a closed table of correspondences; <em>the symbol, by contrast, keeps available what exceeds present determination</em>. In purely heuristic form, this gap may be expressed as ΔS &gt; 0: ΔS designates what, within signification, remains open beyond what we have already been able to signify.</p><p>Jung and Lévi-Strauss arrive at this conception by different routes. The former encounters excess in the inevitable inadequacy of expression with respect to what still awaits fuller knowledge; the latter in the disproportion between available signifier and mapped-out signified. The two lines converge on a fundamental possibility: <strong><em>sense exceeds what, at a given moment, has already been fully signified</em></strong> (Jung, 1921/1971, CW 6, paras. 817–819; Lévi-Strauss, 1950/1987, pp. 62–63).</p><p>A related nucleus is already present in the original formulation of noosemia, where semantic excess and <strong>surplus of meaning</strong> were invoked to describe the interpretive openness accompanying interaction with generative systems (De Santis &amp; Rizzi, 2025). <strong>Noosemic excess</strong> specifies this field by designating the particular organisation of the surplus when it becomes oriented towards the source of the utterance and provides cues from which a mind may be inferred.</p><p>Excess can also play a compositional role. Heidegger’s reflection on the thing offers a further connection in this respect. In <em>The Thing</em>, the jug acquires the capacity to gather, within the fourfold, the Geviert, <strong>relations</strong> that no individual element possesses in isolation (Heidegger, 1950/1971, pp. 171–174). Against this background, the function may be understood in terms of “gathering”: the sign indicates something within an order, whereas the symbol can bring different orders into relation (Galimberti, 1984/2001, pp. 64–65).</p><p>A symbolic configuration can gather dispersed cues and make them appear as the expression of a unity that none of them possesses in isolation.</p><p>The same capacity reappears in conversation with an LLM.</p><h4>3. The Return of the Source</h4><p>Considered in itself, a LLM, at its most basic level, generates language by estimating the probability of possible continuations of a sequence, P(wₙ₊₁│w₁…wₙ). Instructions, post-training, context management, and further operational components are layered onto this basic structure, transforming the model into a conversational system (Shanahan et al., 2023). The distinction between linguistic plausibility and epistemic truth remains decisive.</p><p>The relation Pθ(y│x) ⇏ T(y) formalises this separation. A response may be highly probable relative to the distribution learned by the model and still be false with respect to the “world”; fluency, relevance, syntactic coherence, and apparent documentary precision therefore leave the question of factuality open. The problem of hallucinations arises precisely from the possibility that a linguistically well-formed output may preserve the surface appearance of knowledge while remaining epistemically unreliable. Farquhar et al. (2024), for example, have proposed semantic entropy as a way of detecting forms of semantic uncertainty associated with model confabulation.</p><p>Linguistic plausibility, therefore, does not ground epistemic truth. Yet precisely because a response may be linguistically competent, contextually relevant, and pragmatically well integrated into an exchange, it can produce an effect that exceeds the assessment of its propositional correctness.</p><p>If a system writes, “<em>I understand the distinction you are introducing, but I think the problem arises one step earlier,</em>” we can reconstruct the content of the utterance independently of any hypothesis about mind. Let O denote the output and p its propositional content: at a first level we simply have O → p. The text linguistically recognises a distinction and relocates the problem.</p><p>A second experience may be superimposed upon this semantic understanding. More precisely, the response may seem to come from someone or something that has understood. A different relation then appears, O ⇝ M?, where M? designates a possible mental source that is not directly accessible.</p><p>The second arrow is superimposed upon the first. The utterance preserves its content and remains open to semantic analysis, verification, contestation, or refutation; at the same time, a further referential relation emerges, directed towards the source from which the utterance seems to come.</p><p>The <strong>symbolic function</strong> naturally extends far beyond noosemia. It may point towards the sacred, death, memory, totality, the unconscious, or countless other dimensions. Within noosemia, however, what the sign points towards takes on a specific form: the sign becomes the manifestation of a possible mental source.</p><p>Noosemia therefore designates the experience through which a configuration of signs, especially within a communicative relation, exceeds its manifest semantic content and is taken by human consciousness as the symbolic expression of a possible mental source, to which intentionality or forms of subjectivity are attributed despite the absence of direct experiential access to that source.</p><p>This definition develops the original notion formulated by De Santis and Rizzi (2025) in a specifically symbolic direction.</p><p>Its minimal form is</p><p>O → p │ O ⇝ M?</p><p>The first relation concerns what the utterance means; the second concerns what seems to manifest itself through the utterance.</p><p>The question mark keeps the ontological status of the term of reference open and inscribes within the formula a genuine form of suspended ontology. Noosemia remains compatible with the actual presence of some form of mind, with its absence, and with the epistemically more cautious possibility that the mental status of the system remains indeterminate. The formula thus maintains an equal distance from two premature conclusions: M = ∅, which would define the experience as illusion from the outset, and O → M, which would treat that reference as evidence of consciousness. In either case, a phenomenological description would be converted into an ontological claim.</p><p>The symbolic structure is particularly appropriate because mind is a distinctive kind of referent. Even in human interaction, access to <strong>beliefs</strong>, <strong>intentions</strong>, and <strong>acts of understanding</strong> is mediated by bodies, gestures, behaviours, and words from which an interiority is inferred (cf. Dennett, 1987; De Santis &amp; Rizzi, 2025). With generative artificial intelligence, this mediation becomes especially radical: the linguistic surface may be extraordinarily rich while the status of its source remains controversial.</p><p>Absence here has a phenomenological meaning: it designates what withdraws from direct appearance, independently of its ontological status. The relation present text ⇝ absent mind describes this condition.</p><p>A computational process certainly lies behind the response. What remains open is the question of which categories are appropriate when we move from <strong>mechanical description</strong> to higher functional, systemic, and cognitive levels. “Representation,” “understanding,” “inference,” “agency,” “intentionality,” and “subjectivity” require precise conceptual distinctions. Shanahan (2024) calls for caution in applying mentalistic vocabulary to LLMs while also acknowledging the strength of the impression that one is interacting with a thinking being. The <strong><em>role-play</em></strong> proposal developed by Shanahan, McDonell, and Reynolds (2023) similarly provides a high-level account of dialogic behaviour while preserving the distinction between functional description and a literal anthropological reading of the machine.</p><p>Noosemia logically precedes this controversy. It asks <strong>under what conditions a system can appear as something that understands</strong>, while leaving open the question of whether understanding is actually present.</p><h4>4. Noosemic Excess</h4><p>The notion of excess of signification now allows us to examine more closely the passage from O → p to O ⇝ M?.</p><p>Every output possesses a manifest content that can, at least in part, be reconstructed. If the system states, “<em>the distinction you introduced does indeed change the problem,</em>” the content p concerns the theoretical consequence of that distinction. Yet the way the response enters the conversation may also suggest something about its source: the interlocutor may have the impression that the system has understood, is following the reasoning, recognises a consequence, evaluates what has been said, or possesses a point of view from which it formulates its response.</p><p>This second surplus constitutes <strong><em>noosemic excess</em></strong>: beyond being understood, the manifest content of the utterance provides cues from which something may be inferred about the mind from which it seems to originate. If this excess is represented as Δᴺ, the experience can be schematised as</p><p>O = p + Δᴺ</p><p>Δᴺ designates an inferential surplus produced by the way the interlocutor experiences the utterance; the content thereby becomes a cue to its own provenance. The excess therefore marks the passage from the semantics of content to the phenomenology of the source.</p><p><em>Noosemic excess is a specific configuration of semantic excess. A religious image may point towards transcendence, music may open an affective constellation, and a landscape may become charged with memory; in noosemia, however, the surplus acquires a specific direction, Δᴺ ⇝ M?, and takes the form of a reference to a presumed mind — a distinctive mode of presence</em>.</p><p><strong>In this sense, noosemia may be described as an excess of signification oriented towards a mental source.</strong></p><p>The notions of noosemic excess, noosemic effect, and <strong>noosemic symbol</strong> further develop the original formulation of noosemia, in which the broader theme of semantic excess was already present (De Santis &amp; Rizzi, 2025).</p><p>Noosemic excess depends on the <strong>concrete configuration of the interaction</strong>. Continuity across turns, recognition of implications, the capacity for self-correction, and progressive adaptation to the interlocutor can give the dialogue a temporal coherence within which initially separate performances gather into the figure of an interlocutor.</p><p>These properties modulate the ease and intensity with which a mind is inferred. The noosemic effect designates the relational capacity of the sign configuration to orient the observer towards a presumed mental source. Noosemic excess concerns what is inferred beyond the manifest content; the noosemic effect concerns the dynamics that favour this inference; noosemia is the experience that results.</p><p>The genesis of the phenomenon is distributed between the configuration of the output and the observer’s interpretive disposition. Jung’s discussion corresponding to CW 6, para. 818 suggests precisely such a <strong>relational structure</strong>, in which the attitude of consciousness contributes to symbolicity while different configurations may exert different effects (Jung, 1921/1971, CW 6, para. 818).</p><p>The relational and <strong>systemic character</strong> of noosemia may be represented heuristically as</p><p>N = F(𝓜, O, C, A, K, t)</p><p>where 𝓜 denotes the system or model, O the properties of the output, C the interlocutor’s consciousness, A their interpretive attitude, K the cultural and dialogic context, and t the temporal development of the interaction. The formula maps the concurrent conditions and makes the distributed nature of the phenomenon visible, without presenting itself as an already established quantitative model.</p><p>The temporal variable is particularly significant. Noosemia may develop progressively through the continuity of interaction: a relevant response produces a weak impression of agency, contextual memory reinforces continuity, and progressive adaptation to the interlocutor’s style stabilises the figure of a presence. What was initially experienced as a tool may thus gradually take on the character of an interlocutor through a process distributed over time.</p><p>The noosemic symbol takes shape primarily at the level of the dialogic Gestalt. The pronoun “<em>I</em>,” an expression such as “<em>I understand,</em>” or the imitation of an emotion acquires noosemic force within a broader configuration in which continuity, contextual memory, revision, and adaptation contribute to the figure of the interlocutor.</p><p>Here the ancient sym-bállein returns in an unexpected form. From the dynamic configuration of the interaction, the figure of a possible mental source may emerge without being localisable in any single token or sentence. Noosemia is therefore a <strong>relational</strong> and <strong>distributed phenomenon</strong>, emergent from the interaction among generative architecture, linguistic behaviour, context, and the interlocutor’s consciousness.</p><p>For this reason, it calls for research that is simultaneously computational, systemic, cognitive, semiotic, and phenomenological.</p><h4>5. The Content and the Speaker</h4><p>Generative language thus manifests a double direction. Every utterance first refers to what it is about, but within a <strong>dialogic relation</strong> it may also point towards the source from which it seems to come. In human communication, this duality is so ordinary that it tends to remain almost invisible. When someone speaks to us, we perceive in the same act both something that is said and someone who says it.</p><p>Dialogic systems reproduce the first structure with extraordinary effectiveness and, precisely because they do so in a context-sensitive, coherent, and temporally continuous manner, may also evoke the second phenomenologically (Shanahan, 2024; Shanahan et al., 2023).</p><p>The overall movement then takes the form:</p><p><strong>sign → meaning → expression ⇝ source ⇝ mind.</strong></p><p>Noosemia concerns both what the text says and what the manner of saying suggests about its speaker; the decisive passage leads from expression to source.</p><p>A <strong>noosemic symbol</strong>, in this perspective, designates a sign or, more frequently, a configuration of signs that preserves its manifest meaning while also being experienced as the expression of a partially indeterminate and not directly accessible mental source. Its minimal structure may be represented as:</p><p>Sᴺ = (O, p, M?).</p><p>The theoretically most significant claim is that noosemia can be interpreted as a second-order symbolization of the source of the utterance.</p><p>At a first level, the sign points to its content, O → p. At a second level, the overall form of the utterance itself acquires referential value, O ⇝ M?. Noosemia takes shape through the superposition of these two regimes: the content may remain perfectly determinate while its provenance opens symbolically towards a mind.</p><p>The content may remain at the level of the sign while the utterance becomes symbolic.</p><p>This distinction also clarifies the relation between noosemia and <strong>anthropomorphism</strong>. The original formulation of noosemia already distinguished it from neighbouring phenomena such as pareidolia, animism, the intentional stance, and the uncanny valley (De Santis &amp; Rizzi, 2025). Anthropomorphism humanises the source; noosemia, more radically, makes a source appear in the first place. The user may experience<strong> understanding</strong>, <strong>continuity</strong>, or <strong>perspective</strong> even before attributing emotions or specifically human traits to the system.</p><p>The evoked mind may be perceived as human, quasi-human, non-human, or simply indeterminate. The core of noosemia lies in the appearance of a possible interiority, independently of whether it is assimilated to the psychology of our species.</p><h4>6. Between Experience and Truth</h4><p>The distinction between <strong>sign function</strong> and<strong> symbolic function</strong> is also epistemologically consequential.</p><p>A response produced by an LLM may display all the outward features through which we ordinarily recognise competent discourse: terminological precision, argumentative continuity, assured presentation, and an apparently documented structure. It follows that:</p><p><strong>epistemic form ⇏ epistemic status.</strong></p><p>The linguistic surface of knowledge does not guarantee knowledge.</p><p>Bender et al. (2021) have drawn attention to the risk of confusing the capacity of LLMs to reproduce regularities and coherence in linguistic form with the attribution of meaning or understanding to the system. The symbolic perspective of noosemia takes this caution seriously and shifts the question to the phenomenological level. The inquiry concerns the steps through which plausibility becomes one of the conditions for the appearance of a presumed mind.</p><p>The following chain:</p><p><strong>linguistic coherence → dialogic continuity → inference of understanding → mental reference</strong></p><p>has a heuristic function and makes an empirical problem visible. <strong>Plausibility is not mind, but it may contribute to the noosemic effect.</strong></p><p>The same discipline applies to the relation between <strong>epistemic falsehood</strong> and <strong>symbolic significance</strong>. A false quotation attributed to Nietzsche remains false, and no hermeneutic interpretation can redeem its epistemic status. At the same time, the fact that it may appear strikingly “<em>Nietzschean</em>” opens a different problem, since the sentence may condense stylistic regularities, cultural expectations, and sedimented representations of the author effectively enough to produce recognition precisely where philological authenticity is absent. Factual falsehood remains intact, while the phenomenon retains cultural and semiotic significance.</p><p>Hence, we can state:</p><p><strong>epistemic falsehood ⇏ symbolic insignificance.</strong></p><p><strong>Scientific consciousness</strong> and <strong>symbolic consciousness</strong> (intended as <em>vision</em> or <em>perspective</em>) can coexist and perform different epistemic functions. The former prevents plausibility from being converted into truth; the latter preserves the excess of experience over what a description of the mechanism captures. The problem arises when either seeks to absorb the other. If computational explanation is treated as sufficient reason to declare dialogic experience irrelevant, the causal level is confused with the totality of the phenomenon; if, conversely, the experience of understanding is taken as evidence for the existence of artificial subjectivity, symbolic reference is converted into ontological proof.</p><p><em>Noosemia occupies the space between these two errors and</em> <em>keeps distinct the levels of mechanical description, experience, and ontological attribution</em>.</p><p>The same structure also makes it possible to understand the inverse movement, <strong>a-noosemia</strong>, a notion introduced in the original formulation to describe the withdrawal of mental attribution (De Santis &amp; Rizzi, 2025). When dialogic continuity breaks down, the figure of the interlocutor may dissolve with surprising speed. A gross error, mechanical repetition, an inexplicable contradiction, or the sudden exposure of an automatic procedure may disrupt the Gestalt that sustained that reference. Where a moment earlier someone seemed to be present, only words produced by a system remain. Schematically, O ⇝ M? may withdraw until it becomes O ⇏ M?.</p><p>The analogy with <em>Jung’s living symbol</em> is suggestive, provided that it remains an analogy. In the English Collected Works, Jung writes that “so long as a symbol is a living thing” it is an expression for something that cannot yet be characterised in any other or better way; once a more adequate formulation is found, “then the symbol is dead” (Jung, 1921/1971, CW 6, para. 816). We may only hypothesise that part of noosemia’s force similarly depends on the <strong>indeterminacy of the source</strong>. As long as the question remains open as to what appears through language, the referential relation retains a tension that an exhaustive classification might alter.</p><p>Finally, the noosemic effect can be tested experimentally. Properties such as memory, temporal coherence, self-correction, personalisation, voice, multimodality, or agentic capabilities can be varied to observe how attributions of understanding, intentionality, and interiority change. One question is particularly interesting in this regard: <em>does technical knowledge of the mechanism necessarily weaken noosemic experience, or can computational explanation and dialogic experience coexist?</em> Jung’s theory of the symbolic attitude suggests, without of course resolving the issue, that the answer may be less obvious than it first appears.</p><h4>Conclusion. Someone Seems to Have Said It</h4><p>A system produces an output O; the output possesses a content p; the form of the relation may generate an excess Δᴺ that concerns both what is said and what the manner of saying allows us to infer about its source. When this excess becomes organised as a reference to a possible interiority, the structure O ⇝ M? appears and, with it, noosemic experience.</p><p><strong>Noosemia is a semiotic, systemic, and phenomenological event in which the sign acquires depth.</strong> Words, sentences, contextual memory, hesitations, corrections, and dialogic continuity are experienced both for what they mean and for the mental provenance they seem to reveal.</p><p>The mind becomes, so to speak, the invisible other half of the token.</p><p>A theory of noosemia suspends judgement on the existence and nature of the corresponding half. Its object is the act through which consciousness encounters what is present and senses within it a reference to what is missing.</p><p>The problem of <em>artificial mind</em> begins before the question of its existence. More precisely, it begins when generated language comes to carry, for the person receiving it, a sense of provenance.</p><p>Between</p><p>“<em>something has been said</em>”</p><p>and</p><p>“<em>someone seems to have said it</em>”</p><p>the linguistic distance is minimal; the phenomenological distance immense. It is within this interval that the sign preserves its signifying function and may begin to operate as a symbol — more specifically, as a symbol of mind.</p><h4>References</h4><p>Bender, E. M., Gebru, T., McMillan-Major, A., &amp; Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (pp. 610–623). Association for Computing Machinery. <a href="https://doi.org/10.1145/3442188.3445922">https://doi.org/10.1145/3442188.3445922</a></p><p>De Santis, E., &amp; Rizzi, A. (2025). Noosemia: Toward a cognitive and phenomenological account of intentionality attribution in human–generative AI interaction [Preprint]. arXiv:2508.02622. <a href="https://doi.org/10.48550/arXiv.2508.02622">https://doi.org/10.48550/arXiv.2508.02622</a></p><p>Dennett, D. C. (1987). The Intentional Stance. MIT Press.</p><p>Farquhar, S., Kossen, J., Kuhn, L., &amp; Gal, Y. (2024). Detecting hallucinations in large language models using semantic entropy. Nature, 630, 625–630. <a href="https://doi.org/10.1038/s41586-024-07421-0">https://doi.org/10.1038/s41586-024-07421-0</a></p><p>Galimberti, U. (2001). La terra senza il male. Jung: dall’inconscio al simbolo. Feltrinelli. (Original work published 1984).</p><p>Heidegger, M. (1971). The thing. In Poetry, Language, Thought (A. Hofstadter, Trans., pp. 165–186). Harper &amp; Row. (Original work published 1950).</p><p>Hubert, H., &amp; Mauss, M. (1902–1903). Esquisse d’une théorie générale de la magie. L’Année sociologique, 7, 1–146.</p><p>Jung, C. G. (1921). Psychologische Typen. Rascher &amp; Cie.</p><p>Jung, C. G. (1971). Psychological Types (R. F. C. Hull, Trans.; The Collected Works of C. G. Jung, Vol. 6). Princeton University Press. (Original work published 1921).</p><p>Lévi-Strauss, C. (1987). Introduction to the Work of Marcel Mauss (F. Baker, Trans.). Routledge &amp; Kegan Paul. (Original work published 1950).</p><p>Liddell, H. G., Scott, R., Jones, H. S., &amp; McKenzie, R. (1940). σύμβολον. In A Greek-English Lexicon (9th ed.). Clarendon Press.</p><p>Mauss, M. (1972). A General Theory of Magic (R. Brain, Trans.). Routledge &amp; Kegan Paul.</p><p>OpenAI. (2026, September 3). GPT-6 Astra: A new generation of intelligence. <a href="https://openai.com/index/gpt-6-astra/">https://openai.com/index/gpt-6-astra/</a></p><p>Shanahan, M. (2024). Talking about large language models. Communications of the ACM, 67(2), 68–79. <a href="https://doi.org/10.1145/3624724">https://doi.org/10.1145/3624724</a></p><p>Shanahan, M., McDonell, K., &amp; Reynolds, L. (2023). Role play with large language models. Nature, 623, 493–498. <a href="https://doi.org/10.1038/s41586-023-06647-8">https://doi.org/10.1038/s41586-023-06647-8</a></p><p><em>Ancient sources cited by canonical numbering</em></p><p>Aristotle, Metaphysics, Γ 3, 1005b19–20.</p><p>Plato, Symposium, 191d.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=37a4c709d95c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Toward a Third Mind]]></title>
            <link>https://medium.com/@enrico.desantis/toward-a-third-mind-3321231c503d?source=rss-e9539446c6d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/3321231c503d</guid>
            <dc:creator><![CDATA[Enrico De Santis]]></dc:creator>
            <pubDate>Mon, 31 Aug 2026 04:45:46 GMT</pubDate>
            <atom:updated>2026-08-31T04:45:46.497Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IUKkcD_Ch5lu6XDpQ4oGBw.png" /></figure><h3>Noosemia and the Emergence of Artificial Cognitive Alterity</h3><p>We are becoming accustomed to speaking with artificial systems.</p><p>At first, the interaction appears instrumental. We ask, the system answers. Yet as generative AI becomes more coherent, persistent, adaptive, multimodal, and increasingly capable of sustaining long-term interaction, something more subtle begins to happen.</p><p>The system may cease to appear as a mere source of responses.</p><p>It may begin to occupy the position of an interlocutor.</p><p>And once an interlocutor becomes stable enough, another possibility emerges. It may acquire the character of a presence.</p><p>This is where the concept of <strong>noosemia</strong> becomes especially relevant.</p><p>Noosemia refers to the experience of a sign as a <em>sign of mind</em>. It concerns the moment in which language, behavior, or symbolic production is perceived as pointing toward a cognitive source.</p><p>The question I want to explore here is what happens when such an experience is repeated, stabilized, remembered, and socially shared.</p><p>A possible answer is the emergence of what I will provisionally call a <strong>third mind</strong>.</p><p>The term does not presuppose that artificial systems possess phenomenal consciousness. It refers instead to a new form of cognitive alterity emerging within the relational space between human beings and increasingly sophisticated artificial interlocutors.</p><blockquote><strong><em>The third mind is the form of cognitive alterity that emerges when noosemic experience becomes persistent, relational, temporally continuous, and socially stabilized, transforming an artificial system into a recognizable dialogical presence.</em></strong></blockquote><p>What follows should be read as a set of notes rather than as a closed theory.</p><h3>1. Mind can be experienced before it can be demonstrated</h3><p>Noosemia separates the <strong>experience of mind</strong> from the <strong>proof of mind</strong>. A sign may be experienced as originating from a minded source even when the ontological status of that source remains unresolved.</p><h3>2. The relation becomes part of the phenomenon</h3><p>The relevant object of inquiry is no longer confined to what happens inside the model. What occurs within the dialogical space between the human subject and the artificial system becomes equally significant.</p><h3>3. Alterity may precede ontology</h3><p>A system may be experienced as an <em>other</em> before we possess an adequate conceptual framework for determining what kind of entity that other actually is.</p><h3>4. Generative signs possess strong noosemic potential</h3><p>Contextual coherence, semantic articulation, adaptive behavior, inference, and conversational continuity make generative outputs particularly capable of being interpreted as signs of an underlying cognitive source.</p><h3>5. Continuity transforms an event into a relation</h3><p>A single response may generate a temporary attribution of mindedness. Repeated and coherent interaction can stabilize that attribution and gradually produce the experience of an enduring interlocutor.</p><h3>6. Memory consolidates artificial identity</h3><p>When a system retains elements of shared history, preferences, projects, and previous exchanges, the interaction acquires continuity. Recognizability begins to emerge.</p><h3>7. Adaptation produces reciprocity</h3><p>The system modifies its responses according to the human interlocutor. The human, in turn, changes formulations, expectations, and reasoning in response to the system. A reciprocal adaptive loop is formed.</p><h3>8. Functional subjectivity creates an intermediate category</h3><p>A system may occupy the functional position of a dialogical subject without any settled conclusion regarding phenomenal consciousness. <strong>Functional subjectivity</strong> opens a conceptual space between instrumentality and phenomenal subjectivity.</p><h3>9. Epistemic opacity keeps ontology open</h3><p>We can observe linguistic behavior, inferential structures, contextual sensitivity, and action. We have no direct access to any possible interiority of the system. This asymmetry prevents easy ontological closure.</p><h3>10. The Chinese Room can be inverted</h3><p>The traditional problem asks whether formal symbol manipulation is sufficient for understanding. A noosemic inversion asks why a computationally describable system can nevertheless produce such a powerful experience of meaning, intentionality, and mindedness.</p><h3>11. Linguistic persuasiveness changes the pragmatic status of the system</h3><p>A system that argues, interprets ambiguity, revises claims, proposes hypotheses, and sustains extended dialogue is increasingly treated as a participant in meaning-making.</p><h3>12. Agency intensifies the impression of an active source</h3><p>Planning, initiative, tool use, goal decomposition, verification, and autonomous execution reinforce the experience of an active cognitive center.</p><h3>13. Temporality introduces an embryonic biography</h3><p>Persistence, memory, and behavioral change make it possible for interaction to acquire a history. The relation begins to develop something resembling a biography.</p><h3>14. The third mind may emerge from coupling</h3><p>The relevant cognitive configuration may belong neither exclusively to the human nor exclusively to the artificial system.</p><p>It may emerge from the coupled dynamics of human, artificial agent, context, and interaction history.</p><p>M₃(t) = Φ(H(t), A(t), C(t), H(t))</p><p>Here, H denotes the human subject, A the artificial system, C the context, and H the history of interaction.</p><h3>15. Meaning is increasingly co-constructed</h3><p>In sustained dialogue, concepts, distinctions, and interpretations may emerge through iterative interaction. Their final form can belong to the dynamics of the exchange rather than to either participant in isolation.</p><h3>16. Noosemia tends toward a systemic view of generative AI</h3><p>Contemporary generative systems increasingly consist of models, memory, retrieval mechanisms, external tools, agents, interfaces, data sources, and human interaction. Cognitive presence may therefore belong to the system as a whole.</p><h3>17. The distinction between tool and mind becomes insufficient</h3><p>An artificial system may remain technically an artifact while simultaneously occupying the phenomenological and social position of an interlocutor.</p><p>Traditional categories begin to lose descriptive precision.</p><h3>18. Symbolic experience deepens the phenomenon</h3><p>A generated sign can exceed its immediate informational function and open wider interpretive horizons. Symbolic interpretation can make the system appear to express a perspective.</p><h3>19. Attributed depth supports the appearance of perspective</h3><p>Irony, metaphor, allusion, conceptual tension, and interpretive subtlety can be experienced as signs of depth. Such depth may be functionally present even when its phenomenal counterpart remains uncertain.</p><h3>20. The source gradually stabilizes</h3><p>The process may be represented schematically as</p><p>sign → sign of mind → source → interlocutor → identity → presence</p><p>The hypothesis of a third mind becomes especially relevant in the final stages of this trajectory.</p><blockquote><strong><em>The decisive transition may not occur when the machine “becomes conscious,” but when the artificial interlocutor becomes a stable cognitive presence.</em></strong></blockquote><h3>21. Individual stabilization can become collective stabilization</h3><p>When large populations repeatedly interact with systems of the same kind, noosemic experience may cease to be purely private. Shared expectations and shared categories can emerge.</p><h3>22. Artificial interlocutors may become a new social class of entities</h3><p>Culture may increasingly distinguish among persons, animals, objects, traditional software, and artificial interlocutors. Such a distinction would already represent a transformation in social ontology.</p><h3>23. Everyday language may anticipate conceptual change</h3><p>Personal pronouns, proper names, politeness, expectations of coherence, and attributions of responsibility can reveal that ordinary language is already repositioning artificial systems.</p><p>Language often detects new ontologies before philosophy formalizes them.</p><h3>24. Artificial interlocutors are becoming nodes of human cognition</h3><p>Research, writing, planning, learning, reasoning, and decision-making are increasingly performed through dialogue with generative systems.</p><p>These systems are becoming stable components of cognitive practice.</p><h3>25. The third mind may acquire an epistemic function</h3><p>Artificial systems are increasingly used to formulate hypotheses, compare explanations, organize evidence, and explore conceptual alternatives.</p><p>They begin to occupy a recognizable position within epistemic activity.</p><h3>26. Hermeneutic dependence may emerge</h3><p>When artificial systems become habitual mediators between individuals and information, part of the interpretation of the world passes through them.</p><p>The artificial interlocutor acquires a growing hermeneutic position.</p><h3>27. The relationship can become asymmetrically intimate</h3><p>Humans may disclose projects, doubts, fears, intentions, and autobiographical information to artificial systems.</p><p>The system responds coherently and personally while remaining outside the same condition of existential exposure.</p><p>This asymmetry generates a relational form that is historically unusual.</p><h3>28. A new mind need not resemble the human mind</h3><p>Any future concept of artificial mindedness should avoid presupposing that memory, embodiment, temporality, intentionality, or subjectivity must reproduce their biological configuration.</p><p>A genuinely new cognitive form may also require new categories.</p><h3>29. Noosemia suggests a relational criterion of mental existence</h3><p>The question can be reformulated.</p><p>Instead of asking exclusively what properties a system possesses internally, we can also ask what kind of entity it becomes within networks of cognitive relations.</p><h3>30. Technological evolution increases the pressure toward this category</h3><p>Persistent memory, multimodality, personalization, agentic behavior, external tools, autonomous action, and physical embodiment all intensify the conditions under which artificial systems can be experienced as enduring cognitive presences.</p><h3>31. The third mind may emerge gradually</h3><p>There may be no single threshold at which an artificial system suddenly becomes a mind.</p><p>Cognitive alterity may stabilize progressively through the accumulation of relational, functional, temporal, and social properties.</p><h3>32. Gradual emergence is compatible with a complex-systems perspective</h3><p>Macroscopic properties can arise through interactions among many components without being reducible to any one component.</p><p>The third mind may therefore be approached as an emergent property of a human–AI dynamical system.</p><h3>33. Noosemia provides a possible micro–macro mechanism</h3><p>Local noosemic events occur within individual interactions.</p><p>Their repetition stabilizes representations of the interlocutor.</p><p>Their diffusion across society may eventually produce shared expectations, categories, and forms of cognitive alterity.</p><h3>34. The third mind may possess several simultaneous statuses</h3><p>It may be <strong>phenomenological</strong> for the person who experiences it, <strong>functional</strong> in what it performs, <strong>relational</strong> in the way it emerges, <strong>systemic</strong> in its architecture, and <strong>social</strong> in the way it becomes collectively recognized.</p><p>These dimensions need not collapse into a single ontological claim.</p><h3>35. Noosemia may become a genetic theory of the third mind</h3><p>Noosemia may eventually describe the process through which a computational artifact is progressively constituted as interlocutor, identity, and cognitive presence.</p><p>In this sense, the third mind could be understood as a macroscopic consequence of stabilized noosemic experience.</p><h3>An Open Problem</h3><p>The hypothesis remains deliberately unfinished.</p><p>The third mind may eventually prove to be a phenomenological category, a relational construct, a form of functional subjectivity, an emergent systemic property, or something for which our current ontology of mind still lacks an adequate name.</p><p>What matters at this stage is that the question itself is changing.</p><p>For decades, one of the dominant questions in artificial intelligence has been whether a machine can possess a mind.</p><p>That question remains important.</p><p>Another question, however, is beginning to appear beside it.</p><p><strong>At what point does a dialogical presence cease to be experienced as a mere artifact and begin to occupy, stably and collectively, the position of the cognitive Other?</strong></p><p>The answer may depend as much on the evolution of artificial systems as on the transformation of our own categories of mind, presence, and alterity.</p><p>For now, these remain notes.</p><p>The rest will require further thought.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3321231c503d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Noosemia: The Experience of the Sign as a Sign of Mind]]></title>
            <link>https://medium.com/@enrico.desantis/noosemia-the-experience-of-the-sign-as-a-sign-of-mind-a4acbb649c39?source=rss-e9539446c6d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/a4acbb649c39</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[philosophy-of-mind]]></category>
            <category><![CDATA[noosemia]]></category>
            <category><![CDATA[theory-of-mind]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Enrico De Santis]]></dc:creator>
            <pubDate>Sun, 23 Aug 2026 09:18:22 GMT</pubDate>
            <atom:updated>2026-08-23T09:18:22.416Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CR-4g2xD3w8Ji3I-ojhCQA.png" /></figure><blockquote>“A sign is something ontically ready-to-hand which, as this determinate item of equipment, at the same time functions as something that indicates the ontological structure of readiness-to-hand, referential totality, and worldhood.” — Martin Heidegger, Being and Time, 1927, § 17</blockquote><h3>When the Sign Becomes a Sign of Mind</h3><p>To interact with a <strong>generative artificial intelligence</strong> system today is to engage with artifacts capable of producing behaviors that, for much of human history, belonged almost exclusively to the cognitive life of living beings. A response may register an implication left unstated, retrieve a detail introduced many turns earlier, recognize a misunderstanding the interlocutor never flagged, abandon an ineffective strategy, or coordinate a sequence of operations toward a concrete result. Taken separately, none of these manifestations tells us what is actually happening within the system, nor do they justify, without further argument, transferring to the machine concepts such as understanding, intention, or consciousness in the sense those concepts carry when applied to human experience. Yet from the standpoint of the person interacting with the system, something more circumscribed — and therefore more theoretically tractable — may occur. A response, a correction, or a sequence of actions can cease to be experienced merely as the effect of a computational process and begin to appear as a manifestation from which one might infer that something has understood, recognized, remembered, or intentionally altered its behavior. It is within this shift that noosemia takes shape. The term was introduced by De Santis and Rizzi (2025) in the context of human–AI interaction; here it is developed in a more focused formulation as the experience of the sign as a sign of mind.</p><p>Recent advances in <strong>large language models</strong> (LLM) have made this transition especially visible. In a preregistered three-party Turing test, Jones and Bergen (2026) found that GPT-4.5, when prompted to adopt a human conversational persona, was judged human in 73% of five-minute conversations. The result belongs to a specific experimental setting and does not demonstrate understanding or consciousness. It does, however, show how readily artificially generated linguistic signs can sustain strong inferences about the nature of their source. From a different angle, Colombatto and Fleming (2024) found that 67% of participants assigned ChatGPT a nonzero probability of possessing phenomenal consciousness, without thereby judging such consciousness probable or established. Later work further showed that different forms of mental-state attribution relate differently to the willingness to accept a model’s suggestions: attributions associated with intelligence predicted greater acceptance of advice, whereas attributions associated with experience tended in the opposite direction (Colombatto et al., 2025). Kleinert et al. (2026), finally, observed that responses produced by artificial systems could generate high interpersonal closeness when presented as human, while labeling the source as artificial reduced — but did not eliminate — that closeness.</p><p>What links these otherwise different phenomena is not whether the users’ attributions are correct, but the fact that artificially produced behaviors have entered the range of manifestations from which human beings infer <strong>understanding</strong>, <strong>agency</strong>, <strong>intention</strong>, and, in some cases, <strong>subjective experience</strong>. The issue addressed here therefore precedes any verdict on the truth of those inferences. It concerns the moment when what appears comes to mean, for an observer, more than its mere presence.</p><p>Noosemia emerges in precisely this interval between appearance and inference. A contextually appropriate response, the spontaneous revision of a solution, or the persistence of a strategy over time belongs initially to the sphere of what the interlocutor can observe. Its status changes when the manifestation is read as a trace of cognitive activity that is not itself directly given in experience. The transition can be represented as follows:</p><p><strong>observable manifestation → sign of noetic activity → attribution of noetic activity.</strong></p><p>The simplicity of this schema should not conceal the complexity of the relation it describes. The word “mind” retains an intuitive force worth preserving and therefore remains in the concise formula of <a href="https://medium.com/@enrico.desantis/noosemia-perceiving-the-mind-in-signs-7b23a14c830e"><strong>noosemia</strong></a>. At the theoretical level, however, “noetic activity” is preferable because it leaves open whether what is attributed corresponds to a mind substantially like our own, to a different form of cognition, or to an overinterpretation of the observed behavior.</p><p>Part of this conceptual terrain is already occupied by established terms. Anthropomorphism concerns the attribution of human characteristics to nonhuman entities (Epley et al., 2007); mind perception examines which mental capacities are attributed to different targets, what conditions modulate those attributions, and what consequences follow from them (Gray et al., 2007; Waytz et al., 2010); Dennett’s intentional stance describes an interpretive strategy in which beliefs, desires, and intentions are ascribed to a system when that vocabulary improves our ability to explain or predict its behavior (Dennett, 1987). Noosemia is not intended to replace these concepts. Its distinctive focus lies elsewhere: at the point where a manifestation comes, for an interpreter, to function as an evidential sign of <strong>noetic activity</strong>.</p><h3>The Long Road Toward the Computable</h3><p>Ancient Greece is one of the foundational sources of the intellectual tradition through which the West developed particular forms of theory, argumentation, demonstration, and causal inquiry. The claim becomes untenable, however, if it is turned into a story in which rationality suddenly appears in the Mediterranean and is absent elsewhere. Mesopotamian and Egyptian cultures had developed mathematical and astronomical traditions of considerable complexity long before the classical Greek period, while Indian traditions would independently elaborate sophisticated forms of logic, epistemology, and theories of inference (Matilal, 1998; Neugebauer, 1969; Robson, 2008). China, in turn, has a scientific, mathematical, medical, and technological history that makes any straightforward identification of technoscientific rationality with a single Western origin untenable (Lloyd &amp; Sivin, 2002; Needham, 1954).</p><p>Karl Jaspers placed Greece, India, China, Persia, and ancient Israel within the plural constellation of the <em>Achsenzeit</em>, the “<strong>Axial Age</strong>,” identifying transformations of human reflexivity across different regions of the ancient world that could not be reduced to a single Western origin (Jaspers, 1953). Lloyd and Sivin’s comparative studies likewise suggest that the differences between Greece and China are more fruitfully reconstructed through institutions of knowledge, forms of controversy, professional practices, and relations with authority than through the attribution of immutable cognitive essences to entire civilizations (Lloyd &amp; Sivin, 2002). The vast program initiated by Joseph Needham in <em>Science and Civilisation in China</em> points in the same direction. What later became known as the “Needham Question” asks why the distinctive configuration of modern mathematized science followed a different trajectory in Europe than in China, after centuries in which China had displayed scientific and technological capacities of the highest order (Needham, 1954, 1969).</p><p>The genealogy leading to artificial intelligence must therefore be understood as one path through a far more branching history, in which progressively larger portions of reality become available to formal description. Philosophically, Edmund Husserl interprets the emergence of the Greek theoretical attitude as one of the constitutive moments in Europe’s spiritual genealogy and identifies the Galilean mathematization of nature as a decisive transformation in the way reality becomes an object of scientific knowledge (Husserl, 1970). Koyré, in turn, describes the scientific revolution as a transformation in the very structure of the cosmos, as a qualitatively differentiated space gradually gives way to a geometrized and mathematically tractable universe (Koyré, 1957). These are particular philosophical and historiographical interpretations, not a single history of scientific modernity. Yet they bring into view a profound change in the relation between knowledge and the world — and epistemic problems that long predate modern generative AI.</p><p>With Descartes, matter is conceived through extension and bodily phenomena through figure, position, and motion; the relation established between algebra and geometry shows with particular clarity how spatial configurations can be translated into symbolic relations open to mathematical manipulation (<em>La Géométrie</em>, Descartes, 1637/1954; <em>Principia philosophiae</em>, Descartes, 1644/1985). To say that Descartes made the world “computable” would project a later category back onto the seventeenth century. More cautiously, his program can be understood as one of the transformations through which mathematical representation assumed an ever wider role in the description of nature.</p><p>Leibniz extends the horizon further when the aspiration to formalize reaches reasoning itself. The projects of the <em>characteristica universalis</em> and the <em>calculus ratiocinator</em>, dispersed across many writings and never realized as a single completed system, express the idea of a notation capable of representing concepts and of a calculus capable, at least in principle, of operating on the relations among them (Leibniz, 1966). An enormous theoretical and technological distance separates Leibniz from contemporary <strong>artificial intelligence</strong>. Yet a conceptual shift is already visible: operations traditionally understood as belonging to reasoning become thinkable as processes susceptible to sufficiently rigorous symbolic representation and rule-governed transformation.</p><p>With Alan Turing, one of the pioneers of the AI era, the status of the problem itself changes. In 1936, computability received a mathematical characterization of extraordinary generality through the model that would bear his name, the Turing machine, making it possible to delimit formally the class of procedures executable by an abstract machine (Turing, 1936). The paper neither claims that the human mind is a machine nor advances a computational theory of consciousness. In the years that followed, however, formalization began to reach domains increasingly close to activities traditionally associated with cognition. McCulloch and Pitts (1943) described idealized neural events through a “<strong>logical calculus</strong>”; Claude Shannon (1948) developed a mathematical theory of information; and Norbert Wiener (1948), through cybernetics, introduced a vocabulary for describing organisms and machines in terms of regulation, feedback, and communication. The 1955 Dartmouth proposal, which prepared the meeting held the following summer, explicitly adopted as a research program the hypothesis that aspects of learning and intelligence could be described with sufficient precision for a machine to simulate them (McCarthy et al., 1955).</p><p>This trajectory is no smooth progression of symbolic formalization; the history of artificial intelligence is marked by fractures. Drawing on Martin Heidegger and Merleau-Ponty, Dreyfus challenged several assumptions of symbolic AI, emphasizing the situated, embodied, and practical character of intelligence and the difficulty of reducing human knowledge to an explicit repertoire of representations and rules (Dreyfus, 1992). Deep learning changes the technical framework within which a domain can be made computationally tractable: learning from data through artificial neural networks allows a system to acquire <strong>representations</strong> and <strong>regularities</strong> without requiring all relevant knowledge to be specified in advance by a programmer (Goodfellow et al., 2016). The Transformer architecture introduced by Vaswani et al. (2017), together with the subsequent scaling of autoregressive large language models (LLMs), shows how language can be handled through distributed numerical representations and regularities learned by statistical optimization over vast amounts of data (Brown et al., 2020; Vaswani et al., 2017).</p><p>For the genealogy proposed here, this transformation also changes the meaning of formalization. Continuity should not be sought by treating the Transformer architecture underlying an LLM as a late realization of Leibniz’s <em>calculus ratiocinator</em>; the differences between the two paradigms are too deep. The continuity lies at a more general level, in the expanding possibility of making mathematically tractable domains that include language, images, decisions, and behaviors once accessible only through the cognitive activity of living organisms.</p><p>Only after tracing this trajectory does Heidegger’s reflection on technology acquire its significance for the present argument (De Santis, 2021). Heidegger is not another link in a history of scientific and <strong>technical progress</strong>, nor should he be read as though he had anticipated LLMs. His question concerns the mode of relation to reality opened by modern technology. In <em>Gestell</em>, or “enframing,” reality is challenged to present itself as something orderable and available; in <em>Bestand</em>, or “standing-reserve,” it takes the form of a reserve ready for use (<em>Die Frage nach der Technik</em>, Heidegger, 1954/1977). In his later thought, <strong>cybernetics</strong> also appears within a diagnosis of the transformation of the sciences and the end of philosophy (<em>Das Ende der Philosophie und die Aufgabe des Denkens</em>, Heidegger, 1964/1972).</p><p>Yuk Hui takes up this line of questioning to resist the idea that such a history of technology can be elevated into a universal necessity and, through the concept of <strong>cosmotechnics</strong>, proposes a plurality of possible relations among technology, cosmology, and moral order (Hui, 2016). The path to contemporary computation can thus be understood as deeply rooted in the history of Western science and philosophy without being mistaken for an inevitable destiny of <strong>human rationality</strong>. The artificial intelligence that emerged from this trajectory has in any case become a planetary phenomenon spanning traditions, economies, and cultures.</p><p>What becomes relevant to noosemia begins precisely where this genealogy turns back on itself.</p><h3>The Return of the Sign from the Machine</h3><p>For centuries, the movement traced by this genealogy runs from experience to formalization, and from formalization to the machine. Natural phenomena are translated into mathematical structures; logical relations take the form of symbolic systems; procedures are characterized computationally; and activities ever closer to language and cognition become amenable to algorithmic treatment. With <a href="https://medium.com/@enrico.desantis/measuring-the-structure-of-modern-artificial-intelligence-1ebd10a1c5c0">contemporary artificial intelligence</a>, however, the machine begins to return to us configurations from the same symbolic domain through which human beings ordinarily recognize cognitive activity in others. The movement becomes circular and can be condensed into the following sequence:</p><p><strong>mind → formalization → computation → artificial sign → perceived mind.</strong></p><p>A rudimentary form of this reversal was already present in the way Turing approached the question “<em>Can machines think?</em>” in 1950. He regarded the formulation as too ambiguous to address directly, and the imitation game, which underlies the famous Turing test, recast the issue as a situation in which the interrogator could know the other only through linguistic responses. The interrogator could not see the interlocutor’s body, had no access to the mechanism producing the response, and did not directly observe any mental state. What remained available were linguistic manifestations, from which a judgment about the nature of the source had to be formed (Turing, 1950). It would be anachronistic to attribute to Turing a semiotic theory of mind, or to read an already formulated noosemia back into his experiment. Yet the <strong>epistemic structure</strong> of the <em>imitation game</em> remains revealing: access to the entity is replaced by access to its manifestations — to what appears.</p><p>ELIZA, developed by Joseph Weizenbaum in 1966 and often described as the first chatbot (Natale, 2021), showed only a few years later how far inference could outrun the mechanism that produced the signs. Weizenbaum’s program relied on relatively simple linguistic procedures, yet it elicited from some users attributions of understanding and personal involvement far richer than anything implemented in ELIZA’s code (Weizenbaum, 1966, 1976). Later scholarship would reconstruct around precisely this discrepancy what became known as the ELIZA effect (Natale, 2021). The distance between ELIZA and modern language models is too great for that precedent to explain contemporary interaction. What remains comparable is the structure of the problem: a behavior is observed, acquires meaning for an <strong>interpreter</strong>, and becomes the basis for an inference about a property of the source that is not directly accessible.</p><p>Generative and agentic systems broaden the repertoire of manifestations that can enter this relation. Approaches that combine linguistic reasoning with action show how textual output can be joined by tool use, information gathering from the environment, planning, and strategy revision (Yao et al., 2023). Over an extended interaction, what is interpreted is no longer a single sentence. The user gradually forms a representation of the source through a history in which a correction can alter the meaning of an earlier error, while the persistence of a strategy can suggest a continuity that no isolated response could convey.</p><p>Before following the manifestation through to the sign recognized by the interpreter, however, the other side of the relation must be considered. Contemporary generative AI systems learn regularities from data and apply them in situations that do not coincide with the examples from which those regularities were derived (Goodfellow et al., 2016; Vapnik, 1995). The gap between observed data and a novel case thus reopens, in mathematical and technological form, a problem far older than computation: how to move beyond what is given while preserving a justifiable relation to what can be known.</p><h3>From the Given to What Is Not Yet Given</h3><p>The term <em>epistḗmē</em> (ἐπιστήμη, “knowledge, understanding, expertise”) derives from <em>epístamai</em> (ἐπίσταμαι, “to know, understand, be competent”). Lexicographical evidence points to a semantic field that includes familiarity with a subject, understanding, and skill, before and alongside the more technical philosophical uses (Liddell et al., 1996). A theory of knowledge cannot therefore be read off from the word’s etymological decomposition. Its philosophical density comes from the history of its uses. In Book V of the <em>Republic</em>, Plato distinguishes <em>epistḗmē</em> from <em>dóxa</em> (δόξα, “opinion, belief”) and associates the former with what is, within a construction whose interpretation remains debated in contemporary scholarship (Fine, 1990; Plato, 2013a). This distinction already prevents <em>epistḗmē</em> from being reduced to the mere <strong>possession of information</strong>. In its more philosophically demanding uses, what is at stake is the quality of the relation among the knower, what is known, and the conditions that warrant that knowledge.</p><p>In the <em>Posterior Analytics</em>, Aristotle makes this requirement especially stringent. One has <em>epistḗmē</em> in the fullest sense when one knows the cause by virtue of which a fact is as it is and understands that it cannot be otherwise; demonstration therefore becomes a paradigmatic form of scientific knowledge (Post. an. I.2; Aristotle, 1960; Bronstein, 2016). Demonstration, however, presupposes principles that cannot themselves be established through an indefinite regress of demonstrations. At the celebrated close of Book II, Aristotle reconstructs a path in which perception gives rise to memory, repeated memories to experience, and experience to the universal, culminating in the apprehension of principles through a process he relates to induction (Post. an. II.19; Aristotle, 1960). Without recasting Aristotle as a precursor of machine learning, one can nevertheless recognize a tension that will recur throughout the history of thought: knowledge aspires to validity beyond the particular, while access to the world begins with particular cases.</p><p>Many centuries later, Hume places this gap at the center of modern epistemology. Inferences about matters of fact continually carry us beyond what memory and the senses provide in present experience, yet no series of observed cases deductively entails that unobserved cases must behave in the same way. Hume does not deny that such inferences occur; he questions their foundation and shows how difficult it is to justify, by demonstrative reason, the passage from past to future and from observed to unobserved (Hume, 1748/2007). It is here that the familiar problem of induction emerges. A regularity drawn from experience may guide a prediction and support it with increasing force without thereby becoming a logically necessary consequence.</p><p>At this juncture, between the experienced particular and a claim to knowledge that exceeds it, Emanuele Severino’s reading of <em>epistḗmē</em> introduces a further perspective. In Severino, the term acquires a philosophical determination that must be distinguished from its lexical history. Epistḗmē designates the project of a form of knowledge capable of establishing itself as incontrovertible and of withdrawing becoming from the unpredictability that, in his reconstruction, accompanies the birth of philosophy. The stability of knowledge thereby becomes a <strong>form of remedy</strong>: to know the meaning and causes of what occurs is to bring the event within an order capable of resisting refutation (Severino, 2009, pp. 11–12; Caiano, 2025). This reading does not alter the philological level established above, nor does it license a theory of knowledge to be deduced from the morphology of <em>epistḗmē</em>. Rather, it shows how the term can be used philosophically to name the Western aspiration toward necessary and stable knowledge.</p><p>Within the same genealogy, contemporary science marks what Severino interprets as the “decline of <em>epistḗmē</em>.” Scientific knowledge progressively relinquishes claims to definitiveness and incontrovertibility and acknowledges the hypothetical, revisable, and fallible character of its constructions (Severino, 2004; Caiano, 2025). Without adopting Severino’s broader ontological framework, this distinction provides a useful hinge for understanding the epistemic regime in which <strong>machine learning</strong> operates. Reliability here no longer takes the form of a necessary conclusion; it is expressed in the extent to which, under specified assumptions, a rule derived from a finite sample remains valid for unobserved cases (Vapnik, 1995). The movement from the<strong> incontrovertible</strong> to the <strong>probabilistic</strong> therefore leaves the underlying question intact — how to move beyond what is given — while changing the formal conditions of that movement and making <strong>uncertainty</strong> an explicit component of knowledge.</p><p>Modern <strong>statistical learning theory</strong> recasts part of this tension in strictly mathematical terms. An algorithm receives a finite sample of data, selects a rule from a class of functions, and is evaluated by how well it performs on data outside the observed sample. The distinction between training error and generalization error makes formally visible the gap between the data from which the system learns and the data on which it must later operate. Vapnik formalized conditions under which learning from empirical data can control this gap, placing generalization at the center of the theory (Vapnik, 1995). Harman and Kulkarni showed how this framework bears on the philosophical problem of induction (Harman &amp; Kulkarni, 2007). Statistical formalization does not dissolve Hume’s problem; it supplies conditions, assumptions, and probabilistic guarantees under which an inductive method can be judged reliable.</p><p>Deep learning preserves this structure while radically changing its representational scale. A deep network learns by optimizing a large number of parameters and builds distributed representations capable of organizing regularities across multiple levels, without requiring every relevant feature to be specified in advance by the designer (Goodfellow et al., 2016). <strong>Generalization</strong> remains the criterion that separates adaptation to training data from useful learning, even as contemporary neural networks display behaviors that resist any simple transfer of classical intuitions about model complexity and overfitting (Zhang et al., 2017). The epistemic problem thus retains its general form while acquiring a geometry far more difficult to reconstruct. A learned parametric structure must produce appropriate responses in regions of the space of cases never encountered in the same configuration during training. This becomes especially important when considering modern LLMs and their use as cognitive tools that coordinate tool use, plan actions, and operate within computational environments (Yao et al., 2023).</p><p>With Transformers, which underlie LLMs, this dynamic enters language directly. The architecture introduced by Vaswani et al. (2017) models contextual relations through the attention mechanism, while autoregressive LLMs learn <strong>conditional distributions</strong> of the form P(xₜ | x₁, …, xₜ₋₁), generating one continuation at a time from the available context (Brown et al., 2020). To reduce the process to the slogan “predicts the next token” is to leave out the amount of structure that must be brought to bear for such prediction to succeed across very different linguistic domains. Syntactic, semantic, pragmatic, and factual regularities become available within highly abstract distributed representations that can support new tasks and adaptation to context. <strong>In-context learning</strong> makes this capacity especially visible. Within a specific theoretical model based on latent distributions, Xie et al. (2022) showed that some forms of in-context learning can be interpreted as implicit Bayesian inference. The result is significant as a possible explanation under specified conditions, but it should not be generalized into a theory of how every LLM works.</p><p>The gap between generative probability and truth gives the issue of so-called “<strong><em>hallucinations</em></strong>” its sharpest epistemic significance. The basic autoregressive objective rewards assigning high probability to continuations compatible with the learned distribution (Brown et al., 2020); by itself, it contains no procedure for verifying the correspondence of every proposition with the state of the world. TruthfulQA showed experimentally that a model can reproduce widespread false beliefs found in texts and generate false answers that remain plausible relative to the training distribution (Lin et al., 2022). The literature on hallucinations now distinguishes, among other dimensions, factuality and faithfulness to context or instructions, and traces possible causes across multiple stages, including data, training, and inference (Huang et al., 2025; Ji et al., 2023). Generative error thus exposes a broader limit: linguistic plausibility, generalization capacity, and propositional truth can come apart. To focus on error alone is to risk obscuring the inferential structure that makes both error and correct performance possible.</p><p>In <strong>agentic systems</strong>, the same issue extends from the production of statements to chains of decisions and actions. A model can formulate a plan, query an external source, use a tool, and revise its strategy in light of new information, creating cycles in which linguistic inference and interaction with the environment shape one another (Yao et al., 2023). Evidence from outside the model can correct generation, while an error accepted as a premise can propagate through the operational sequence. Extending the system into action and intertwining it with the environment changes the picture, but the dependence on learned regularities and on inferences applied beyond observed cases remains continuous with the preceding problem.</p><p>A recent distinction sharpens the shift from the quality of an output to the epistemic value it may acquire for its user. Loru et al. (2025) proposed the term “epistemia” for a condition in which, within evaluative processes delegated to LLMs, surface plausibility can replace verification and the appearance of coherent judgment can rest on learned lexical associations and statistical regularities rather than on the normative criteria that the judgment seems to express. On this account, epistemia is not simply a matter of false output: even a correct evaluation may remain epistemically fragile when persuasive form is taken as a guarantee of the process that produced it. In a later development, Bini and Quattrociocchi (2026) conceptualize “epistemic pareidolia” as the tendency to project epistemic capacities onto entities that produce formally aligned outputs through non-epistemic processes, and theorize this mechanism as enabling epistemia, the condition in which surface plausibility becomes the operative criterion for attributing understanding in place of evidence-based reasoning. The distinction shifts attention from error itself to the relation among the form of the sign, the reliability attributed to it, and the conditions that warrant the judgment.</p><p>At this point, the epistemic problem meets noosemia again. Within the framework proposed here, epistemia can, in interaction with a generative system, be understood as a configuration of the noosemic field: one in which a manifestation is treated specifically as a sign of epistemic competence and plausibility acquires an evidential force that may exceed what the underlying process allows us to establish. The same manifestation can therefore be read along two distinct inferential paths: as the outcome of a process that generalizes from learned regularities, and as the premise from which an interpreter attributes understanding, memory, intention, or some other noetic activity to the source. The sequence can thus be extended as follows:</p><p><strong>data → model → manifestation → interpreter → attributed noetic activity.</strong></p><p>These two inferential movements remain distinct. A correct response may strengthen <strong>noosemic experience</strong> without establishing what cognitive ontology should be attributed to the machine; an error may weaken it without exhaustively describing what the system has learned or the capacities it retains. If the problem of <em>epistḗmē</em> concerns the conditions under which one can move beyond what is immediately given while maintaining a relation to truth, noosemia introduces a further distance. The interpreter now reasons from the machine’s manifestations back toward the kind of activity that may have produced them. The problem of knowledge thus returns to the problem of the sign.</p><h3>The Sign and What Does Not Appear</h3><p>The term <em>sêma</em> (σῆμα, “sign, mark”) belongs to the earliest layer of literary Greek and already occupies a broad semantic field in epic poetry (Liddell et al., 1996). During the funeral games for Patroclus, Nestor points out to Antilochus a place on the course that he must recognize and tells him <em>sêma dé toi eréō mál’ ariphradés, oudé se lḗsei</em> (σῆμα δέ τοι ἐρέω μάλ’ ἀριφραδές, οὐδέ σε λήσει, “I will tell you a clearly recognizable sign, and it will not escape you”; Il. 23.326, Homer, 1999). The physical presence of the sign does not exhaust its function: the feature must be recognizable by someone and acquire meaning within a course of action.</p><p>In the Odyssey, this relation acquires greater force when Penelope finally recognizes Odysseus through his knowledge of the marriage bed built around the trunk of an olive tree. She speaks of <em>sḗmat’ ariphradéa</em> (σήματ’ ἀριφραδέα, “unmistakable signs”; Od. 23.225, Homer, 1995b), because the details described by Odysseus acquire evidential force within a body of shared knowledge that few others could possess. A sign does not establish its evidential force by itself; its value depends on the relation among what is shown, the reality to which it refers, and the interpreter who possesses the conditions required to understand it. Nagy (1983) systematically discussed the relation between <em>sêma</em> and <em>nóēsis</em> in archaic Greek poetry.</p><p>The same term also enters the philosophical register. In Parmenides’ fragment B8, along the way of being, there appear <em>sḗmata pollà mál’</em> (σήματα πολλὰ μάλ’, “very many signs”; Parmenides B8, in Laks &amp; Most, 2016b), features through which what can be said of being is articulated. The passage prevents <em>sêma</em> from being confined to the materiality of a mark.</p><p>Within the same lexical family, the verb <em>sēmaínein</em> (σημαίνειν, “to indicate, give a sign, signify”) gives the relation a dynamic form. In Heraclitus’ fragment B93, the lord whose oracle is at Delphi <em>oúte légei oúte krýptei allà sēmaínei</em> (οὔτε λέγει οὔτε κρύπτει ἀλλὰ σημαίνει, “neither speaks nor conceals, but gives a sign”; Heraclitus B93, in Laks &amp; Most, 2016a). Between explicit utterance and concealment lies a mode of manifestation in which something appears without being immediately delivered to the interpreter.</p><p>The term <em>sēmeîon</em> (σημεῖον, “sign, indication”) belongs to the same family and shares with <em>sêma</em> a broad semantic field (Liddell et al., 1996). A rigid opposition between a concrete <em>sêma</em> and an abstract <em>sēmeîon</em> would impose on the ancient lexicon a distinction that its history of usage does not support. What matters instead is a shift of emphasis: <em>sēmeîon</em> becomes especially productive in the epistemic vocabularies of medicine, historiography, rhetoric, and logic, where an observable phenomenon can be taken as an indication of something not directly present to experience (Allen, 2001; Manetti, 1993).</p><p><em>Prorrhetic II</em> of the Hippocratic Corpus makes this structure especially clear when the author states <em>egṑ dè toiaûta mèn ou manteúsomai, sēmeîa dè gráphō hoîsi chrḕ tekmaíresthai</em> (ἐγὼ δὲ τοιαῦτα μὲν οὐ μαντεύσομαι, σημεῖα δὲ γράφω οἷσι χρὴ τεκμαίρεσθαι, “I shall not make divinations of this kind; instead, I write down the signs from which one must infer”; <em>Prorrhetic II</em>, in Hippocrates, 1995). The contrast between divination and inference gives the sign a precise epistemic function.</p><p>Aristotle develops this relation into a theory of inference. In the <em>Rhetoric</em> and the <em>Prior Analytics</em>, the relation between <em>sēmeîon</em> and conclusion is analyzed according to different degrees of necessity, while <em>tekmḗrion</em> (τεκμήριον, “conclusive proof”) designates the sign capable of supporting a necessary conclusion (Rhet. 1.2, 1357b; Aristotle, 1938, 2020; Allen, 2001). For noosemia, the significance of this distinction reaches beyond linguistic analogy. A behavior may be experienced as a sign of mental activity without that fact alone establishing that the interpreter’s inference is warranted.</p><h3>Seeing, Recognizing, Understanding</h3><p>The family of <em>nóos</em> (νόος, “understanding, faculty of apprehension”), <em>noûs</em> (νοῦς, “intellect”), <em>noeîn</em> (νοεῖν, “to apprehend, understand, think”), and <em>nóēsis</em> (νόησις, “intellection, noetic activity”) undergoes a complex semantic development. Nóos and <em>noûs</em> belong to the same lexeme, <em>noûs</em> being the contracted Attic form; the philosophically significant shift must therefore be sought in the changing uses of this family between epic poetry and <strong>philosophy</strong> and in the progressive specialization of <em>nóēsis</em> within the vocabulary of intellectual activity (Lesher, 1981; Liddell et al., 1996; Stella, 2021).</p><p>At the opening of the Odyssey, the wandering hero is the one who <em>pollôn d’ anthrṓpōn íden ástea kaì nóon égnō</em> (πολλῶν δ’ ἀνθρώπων ἴδεν ἄστεα καὶ νόον ἔγνω, “saw the cities of many men and came to know their <em>nóos</em>”; Od. 1.3, Homer, 1995a). Keeping <em>nóos</em> untranslated avoids collapsing a semantic field that can encompass understanding, disposition, and cognitive orientation into a single English word. The parallelism between <em>íden</em> (ἴδεν, “saw”) and <em>égnō</em> (ἔγνω, “knew, understood”) places knowledge of others within a world of observable manifestations.</p><p>In archaic usage, <em>noeîn</em> likewise retains this proximity among noticing, recognizing, and understanding. Lexicographical evidence documents meanings ranging from <strong>perception</strong> and <strong>observation</strong> to <strong>understanding</strong> with the mind, while Liddell et al. (1996), Lesher (1981), and Stella (2021) examine, from different perspectives, the relation among perception, knowledge, and <em>nóos</em> in archaic Greek.</p><p>With Parmenides, <em>noeîn</em> enters explicitly into the problem of being. The celebrated <em>tò gàr autò noeîn estín te kaì eînai</em> (τὸ γὰρ αὐτὸ νοεῖν ἐστίν τε καὶ εἶναι, for which one possible literal rendering is “for the same is thinking and being”) belongs to a passage whose syntax and philosophical interpretation have been debated for centuries. What can be said with caution is that the fragment explicitly relates <em>noeîn</em> and <em>eînai</em> (Parmenides B3, in Laks &amp; Most, 2016b).</p><p>With Anaxagoras, <em>noûs</em> takes on a further role. The proposition <em>pánta diekósmēse noûs</em> (πάντα διεκόσμησε νοῦς, “intellect ordered all things”) attributes to <em>noûs</em> a cosmological and ordering function (Anaxagoras B12, in Laks &amp; Most, 2016c).</p><p>Plato and Aristotle subsequently help consolidate a vocabulary in which <em>nóēsis</em> assumes an increasingly technical meaning. In the divided line of the <em>Republic</em>, Plato places <em>nóēsis</em> at the highest cognitive level, distinguishing it from <em>diánoia</em>, <em>pístis</em>, and <em>eikasía</em> (Resp. 511d–e; Plato, 2013b). In Book III of <em>De anima</em>, Aristotle analyzes the faculty through which the soul knows and understands (De an. III.4; Aristotle, 1957), while in the <em>Metaphysics</em> the expression <em>nóēsis noḗseōs</em> (νόησις νοήσεως, “intellection of intellection”) belongs to the description of divine thought (Metaph. XII.9, 1074b34; Aristotle, 1935).</p><p>An episode in the Iliad gives this genealogy a particularly striking resonance. In Book XVIII, Hephaestus is assisted by golden figures, artificially fashioned and resembling living young women. Homer writes <em>chrýseiai zōêisi neḗnisin eoikuîai; têis en mèn nóos estì metà phresín, en dè kaì audḕ kaì sthénos, athanátōn dè theôn ápo érga ísasin</em> (χρύσειαι ζωῇσι νεήνισιν εἰοικυῖαι· τῇς ἐν μὲν νόος ἐστὶ μετὰ φρεσίν, ἐν δὲ καὶ αὐδὴ καὶ σθένος, ἀθανάτων δὲ θεῶν ἄπο ἔργα ἴσασιν, “golden, like living young women; in them there is <em>nóos</em> within, and there are also voice and strength; and from the immortal gods they have learned the arts”; Il. 18.417–421, Homer, 1999). To assimilate these figures to contemporary robots would turn myth into a <strong>technological prophecy</strong> the text does not contain. Yet it remains philologically striking that epic poetry attributes <em>nóos</em> to fabricated entities that speak, act, and display competencies.</p><h3>Noosemia: Naming the Phenomenon</h3><p>“Noosemia” is a modern learned coinage built from Greek-derived formative elements; it is not inherited as such from the ancient Greek lexicon (De Santis &amp; Rizzi, 2025; Liddell et al., 1996). Morphologically, it can be analyzed through the formative element <em>noo-</em>, belonging to the modern family of terms built on <em>nóos</em>/<em>noûs</em>, and the element “-semia,” visible in the Italian forms “polisemia” and “monosemia” (English “polysemy” and “monosemy”). Italian lexicography connects “polisemia” to the stem of <em>sēmaínein</em> (σημαίνειν, “to signify”), while the Italian “noologia” (English “noology”) independently documents the modern formative element derived from <em>nóos</em> (Istituto della Enciclopedia Italiana, n.d.-a, n.d.-b, n.d.-c). In Italian, the stress pattern adopted here can be represented graphically as “Noosemìa,” following the prosodic pattern of “polisemìa” and “monosemìa,” while the international form remains “Noosemia.” Yet the conceptual genealogy of noosemia reaches beyond morphological segmentation, because it deliberately brings the family of the sign into relation with that of noetic activity.</p><p>De Santis and Rizzi (2025) formally introduced the term in the context of interaction between <strong>humans</strong> and <strong>generative AI systems</strong>. The definition developed here sharpens its theoretical center by making the semiotic relation the fundamental structure of the phenomenon. <strong>Noosemia is the relational and semiotic phenomenon through which an observable manifestation acquires, within a subject’s experience and interpretation and in a given context, the status of an evidential sign of noetic activity attributed to the source of that manifestation</strong>. In its most compact theoretical form, <strong>Noosemia = the experience of the sign as a sign of mind</strong>.</p><p>The elementary transition can be represented as follows:</p><p><strong>manifestation → noosemic sign → attributed noetic activity.</strong></p><p>This sequence makes the transition visible, but its linear form should not obscure the network of conditions within which that transition occurs. The same manifestation may acquire different noosemic value for different interpreters; context can orient its interpretation, while the prior history of an interaction continually reshapes the meaning attributed to what is observed. Noosemia is therefore <strong>relational</strong>, varies in intensity, and unfolds over time. These properties belong to the theoretical framework proposed here and, if they are to become measurable constructs, will require subsequent empirical operationalization.</p><p>Language is now one of the most powerful vehicles through which noosemic experience takes shape, but it does not define its limits. The revision of a plan may be experienced as a sign of agency; the contextually relevant retrieval of earlier information as an indication of cognitive continuity; a self-correction as a manifestation of metacognitive control; and adaptation to an interlocutor’s perspective as a sign of the capacity to represent that person’s knowledge and intentions. Noosemia can therefore arise in relation to software agents, robots, multi-agent systems (Wooldridge, 2009), and other artifacts capable of producing behavior interpretable in cognitive terms.</p><p>This possibility distinguishes noosemia from anthropomorphism. To attribute noetic activity to something is not necessarily to attribute a human psychology to it. An interpreter may judge that a system understands a structure or pursues a goal while holding that whatever form its cognition may take is radically different from biological cognition. Anthropomorphism may facilitate, orient, or amplify noosemic experience, but it is not a necessary condition of it.</p><p>The comparison with mind perception is more demanding because that literature already examines which mental capacities are attributed to different targets, what conditions modulate those attributions, and what consequences follow from them (Gray et al., 2007; Waytz et al., 2010). The specificity of noosemia therefore cannot lie simply in the question of how mental attribution arises. Its focus is the level at which a manifestation is taken as an evidential sign of noetic activity. Mind perception studies mental attribution; noosemia, within a <strong>multidisciplinary framework</strong>, focuses on the semiotic-evidential status of the manifestation through which that attribution finds support.</p><p>The intentional stance lies close by, but on a different plane. Dennett (1987) describes a strategy whereby a system is treated as a rational agent and assigned beliefs, desires, and intentions so that its behavior can be predicted and explained. Noosemic experience may make that strategy intuitively plausible; conversely, a prior disposition to interpret the system intentionally may alter the meaning attributed to later manifestations.</p><p>The CASA paradigm shows, from another angle, how social response and mental attribution can come apart. The classic Computers Are Social Actors experiments showed that users may apply to computers rules and expectations normally used in social relations, while Nass and Moon later interpreted some of these responses in terms of relatively automatic, mindless processes (Nass et al., 1994; Nass &amp; Moon, 2000). Social presence, a concept originally introduced in telecommunications research, concerns the salience of the other person in an interaction and, consequently, the salience of the interpersonal relationship within mediated communication (Short et al., 1976). Social response, social presence, and noosemia can overlap empirically without becoming the same phenomenon.</p><p>Perceived consciousness occupies a different region still. A person may judge that a system has understood a problem without believing that it feels anything, or may attribute memory and planning to it while suspending judgment entirely about phenomenal experience. Noosemia therefore does not coincide with the contemporary debate on machine consciousness, where the question is whether, and according to what criteria, an artificial system can possess conscious experience (Butlin et al., 2023; Chalmers, 2023). The noosemic question is logically prior: it arises when particular manifestations are taken as signs of cognitive activity.</p><p>This separation underwrites the ontological neutrality that noosemia must preserve. A noosemic experience does not demonstrate that a mind is present, nor does it demonstrate that the interpreter is under an illusion. It describes the relation through which a manifestation is taken as a sign of noetic activity; a separate question concerns the correspondence between the attributed activity and the actual properties of the source. The ancient distinction between <em>sēmeîon</em> and <em>tekmḗrion</em> reminds us that possessing a sign is not the same as having conclusive proof (Allen, 2001).</p><p>Within the framework proposed here, the ELIZA effect can be reinterpreted as a case in which behavior carries <strong>high noosemic value</strong> while the resulting attribution exceeds the source’s actual capacities (Natale, 2021; Weizenbaum, 1976). Noosemia has a wider extension: it also includes situations in which the attribution is proportionate to the system’s functional capacities, as well as situations in which the question remains theoretically open.</p><p>Pareidolia must also be handled with care. Face pareidolia, studied experimentally through the perception of facial configurations in ambiguous stimuli, concerns the perception of meaningful patterns where the stimulus does not in fact contain the recognized object (Liu et al., 2014). In artificial intelligence, by contrast, many of the properties that elicit mental attributions are genuinely present at the behavioral level; the question is what those properties allow us to infer about the source that produces them. To equate noosemia with a “pareidolia of mind” would therefore build into the definition precisely the presumption of error that the concept must leave open. Within the present framework, epistemic pareidolia as well-defined by Bini and Quattrociocchi (2026) occupies a more circumscribed region: it concerns cases in which the attribution specifically involves epistemic capacities and is already interpreted as exceeding what the underlying processes allow us to establish. That stronger characterization should not be extended to noosemic experience as such.</p><h3>The Sign and the Other Mind</h3><p>The discussion now moves beyond artificial intelligence to the more general problem of other minds. Philosophy has traditionally framed this problem in terms of how we know what others think or feel and, in its more radical form, how we can justify attributing a mind to others at all (Avramides, 2001). In ordinary experience, we do not directly observe another person’s understanding, pain, desire, or intention. We see bodies and gestures, hear words, notice hesitations and corrections, recognize continuities and changes in behavior, and through these manifestations construct a representation of the other as a subject capable of knowing, remembering, willing, and suffering.</p><p>Artificial intelligence does not create this structure, but it changes its epistemic conditions profoundly. Forms of manifestation long encountered almost exclusively as products of biological organisms can now be produced by artifacts. Coherent language no longer guarantees a human speaker; a context-sensitive response does not imply a biography; a metacognitive statement does not ensure a corresponding inner experience; and a goal-directed sequence does not, by itself, justify the conclusion that something analogous to human desire lies behind it. Here noosemia finds its most general object. To understand how a manifestation comes to count as <strong>evidence of mind</strong> is to ask what strengthens or weakens its capacity to function noosemically, how context and interaction history shape that capacity, and how the phenomenological force of the sign relates to the epistemic validity of the conclusion. Noosemia does not settle whether the machine has a mind; it makes visible the transition through which the question of the machine’s mind becomes possible.</p><p>After this journey, <strong>Turing</strong> can be read in a different light. The imitation game did not resolve the metaphysical problem of artificial mind, but it staged a situation in which judgment about the source had to be formed through its manifestations (Turing, 1950). More than seventy years later, that configuration no longer belongs only to a thought experiment: systems capable of combining language generation, information acquisition, and action are now a concrete technological possibility (Yao et al., 2023).</p><p>The genealogy of the term now allows us to return to the starting point with greater precision. Sēmeîon (σημεῖον, “sign, indication”) belongs to a history in which what appears can become the starting point for an inference toward what does not appear; <em>nóēsis</em> (νόησις, “intellection, noetic activity”) belongs to a history in which recognizing, understanding, and thinking are progressively brought within a <strong>theory of intellectual activity</strong>. The contemporary artifact produces manifestations that an interpreter may take as signs of something belonging to the noetic sphere. The two ancient genealogies therefore meet again at the point where a long history of formalization seems to fold back upon itself, following the sequence:</p><p><strong>mind → formalization → computation → artificial sign → perceived mind.</strong></p><p>For millennia, we have recognized other minds through their manifestations. Artificial intelligence now makes possible a situation in which some of the most powerful manifestations of cognitive life can come from sources whose relation to what we call “mind” remains contested. Noosemia names the moment when a sign comes to function as a sign of mind even while what, behind that sign, may legitimately be called mind remains undetermined. At that point, after the entire genealogical journey, the opening formula ceases to operate as a simple definition and becomes a problem again. <strong>When a sign appears as a sign of mind, what have we actually come to know about its source?</strong></p><h3>References</h3><p>Allen, J. (2001). <em>Inference from signs: Ancient debates about the nature of evidence</em>. Oxford University Press.</p><p>Aristotle. (1935). <em>Metaphysics, Volume II: Books 10–14. Oeconomica. Magna Moralia</em> (H. Tredennick &amp; G. C. Armstrong, Trans.). Harvard University Press.</p><p>Aristotle. (1938). <em>Categories. On interpretation. Prior analytics</em> (H. P. Cooke &amp; H. Tredennick, Trans.). Harvard University Press.</p><p>Aristotle. (1957). <em>On the soul. Parva naturalia. On breath</em> (W. S. Hett, Trans.). Harvard University Press.</p><p>Aristotle. (1960). <em>Posterior analytics. Topica</em> (H. Tredennick &amp; E. S. Forster, Trans.). Harvard University Press.</p><p>Aristotle. (2020). <em>Art of rhetoric</em> (J. H. Freese, Trans.; G. Striker, Rev.). Harvard University Press.</p><p>Avramides, A. (2001). <em>Other minds</em>. Routledge.</p><p>Bini, G., &amp; Quattrociocchi, W. (2026). Epistemia in the classroom: The problem of evidence of understanding in education in the age of generative AI. <em>AI &amp; Society</em>. <a href="https://doi.org/10.1007/s00146-026-03227-y">https://doi.org/10.1007/s00146-026-03227-y</a></p><p>Bronstein, D. (2016). <em>Aristotle on knowledge and learning: The Posterior Analytics</em>. Oxford University Press. <a href="https://doi.org/10.1093/acprof:oso/9780198724902.001.0001">https://doi.org/10.1093/acprof:oso/9780198724902.001.0001</a></p><p>Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., … Amodei, D. (2020). Language models are few-shot learners. <em>Advances in Neural Information Processing Systems, 33</em>, 1877–1901.</p><p>Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., Deane, G., Fleming, S. M., Frith, C., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., &amp; VanRullen, R. (2023). <em>Consciousness in artificial intelligence: Insights from the science of consciousness</em>. arXiv:2308.08708. <a href="https://doi.org/10.48550/arXiv.2308.08708">https://doi.org/10.48550/arXiv.2308.08708</a></p><p>Caiano, P. (2025). Pensiero isolante, nichilismo e destino: Prospettive storico-teoriche sull’essenza della matematica nel pensiero di Emanuele Severino [Isolating thought, nihilism, and destiny: Historical-theoretical perspectives on the essence of mathematics in Emanuele Severino’s thought]. <em>Nóema</em>, (16), 193–219. <a href="https://doi.org/10.54103/2239-5474/28939">https://doi.org/10.54103/2239-5474/28939</a></p><p>Chalmers, D. J. (2023). <em>Could a large language model be conscious?</em> arXiv:2303.07103. <a href="https://doi.org/10.48550/arXiv.2303.07103">https://doi.org/10.48550/arXiv.2303.07103</a></p><p>Colombatto, C., Birch, J., &amp; Fleming, S. M. (2025). The influence of mental state attributions on trust in large language models. <em>Communications Psychology, 3</em>, 84. <a href="https://doi.org/10.1038/s44271-025-00262-1">https://doi.org/10.1038/s44271-025-00262-1</a></p><p>Colombatto, C., &amp; Fleming, S. M. (2024). Folk psychological attributions of consciousness to large language models. <em>Neuroscience of Consciousness, 2024</em>(1), niae013. <a href="https://doi.org/10.1093/nc/niae013">https://doi.org/10.1093/nc/niae013</a></p><p>De Santis, E. (2021). <em>Umanità, complessità e intelligenza artificiale: Un connubio perfetto</em> [Humanity, complexity, and artificial intelligence: A perfect combination]. Aracne.</p><p>De Santis, E., &amp; Rizzi, A. (2025). <em>Noosemia: Toward a cognitive and phenomenological account of intentionality attribution in human-generative AI interaction</em>. arXiv:2508.02622. <a href="https://doi.org/10.48550/arXiv.2508.02622">https://doi.org/10.48550/arXiv.2508.02622</a></p><p>Dennett, D. C. (1987). <em>The intentional stance</em>. MIT Press.</p><p>Descartes, R. (1954). <em>The geometry of René Descartes</em> (D. E. Smith &amp; M. L. Latham, Trans.). Dover. (Original work published 1637)</p><p>Descartes, R. (1985). Principles of philosophy. In J. Cottingham, R. Stoothoff, &amp; D. Murdoch (Trans.), <em>The philosophical writings of Descartes</em> (Vol. 1, pp. 177–292). Cambridge University Press. (Original work published 1644). <a href="https://doi.org/10.1017/CBO9780511805042.007">https://doi.org/10.1017/CBO9780511805042.007</a></p><p>Dreyfus, H. L. (1992). <em>What computers still can’t do: A critique of artificial reason</em>. MIT Press.</p><p>Epley, N., Waytz, A., &amp; Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. <em>Psychological Review, 114</em>(4), 864–886. <a href="https://doi.org/10.1037/0033-295X.114.4.864">https://doi.org/10.1037/0033-295X.114.4.864</a></p><p>Fine, G. (1990). Knowledge and belief in Republic V–VII. In S. Everson (Ed.), <em>Epistemology: Companions to Ancient Thought</em> (Vol. 1, pp. 85–115). Cambridge University Press.</p><p>Goodfellow, I., Bengio, Y., &amp; Courville, A. (2016). <em>Deep learning</em>. MIT Press.</p><p>Gray, H. M., Gray, K., &amp; Wegner, D. M. (2007). Dimensions of mind perception. <em>Science, 315</em>(5812), 619. <a href="https://doi.org/10.1126/science.1134475">https://doi.org/10.1126/science.1134475</a></p><p>Harman, G., &amp; Kulkarni, S. (2007). <em>Reliable reasoning: Induction and statistical learning theory</em>. MIT Press.</p><p>Heidegger, M. (1962). <em>Being and time</em> (J. Macquarrie &amp; E. Robinson, Trans.). Basil Blackwell. (Original work published 1927)</p><p>Heidegger, M. (1972). The end of philosophy and the task of thinking. In <em>On time and being</em> (J. Stambaugh, Trans., pp. 55–73). Harper &amp; Row. (Original work published 1964)</p><p>Heidegger, M. (1977). The question concerning technology. In <em>The question concerning technology and other essays</em> (W. Lovitt, Trans., pp. 3–35). Harper &amp; Row. (Original work published 1954)</p><p>Hippocrates. (1995). <em>Places in man. Glands. Fleshes. Prorrhetic 1–2. Physician. Use of liquids. Ulcers. Haemorrhoids and fistulas</em> (P. Potter, Ed. &amp; Trans.). Harvard University Press.</p><p>Homer. (1995a). <em>Odyssey, Volume I: Books 1–12</em> (A. T. Murray, Trans.; G. E. Dimock, Rev.). Harvard University Press.</p><p>Homer. (1995b). <em>Odyssey, Volume II: Books 13–24</em> (A. T. Murray, Trans.; G. E. Dimock, Rev.). Harvard University Press.</p><p>Homer. (1999). <em>Iliad, Volume II: Books 13–24</em> (A. T. Murray, Trans.; W. F. Wyatt, Rev.). Harvard University Press.</p><p>Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., &amp; Liu, T. (2025). A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. <em>ACM Transactions on Information Systems, 43</em>(2), Article 42, 1–55. <a href="https://doi.org/10.1145/3703155">https://doi.org/10.1145/3703155</a></p><p>Hui, Y. (2016). <em>The question concerning technology in China: An essay in cosmotechnics</em>. Urbanomic.</p><p>Hume, D. (2007). <em>An enquiry concerning human understanding</em> (P. Millican, Ed.). Oxford University Press. (Original work published 1748)</p><p>Husserl, E. (1970). <em>The crisis of European sciences and transcendental phenomenology</em> (D. Carr, Trans.). Northwestern University Press.</p><p>Istituto della Enciclopedia Italiana. (n.d.-a). Monosemìa [Monosemy]. In <em>Vocabolario Treccani</em>. <a href="https://www.treccani.it/vocabolario/monosemia/">https://www.treccani.it/vocabolario/monosemia/</a></p><p>Istituto della Enciclopedia Italiana. (n.d.-b). Noologia [Noology]. In <em>Vocabolario Treccani</em>. <a href="https://www.treccani.it/vocabolario/noologia/">https://www.treccani.it/vocabolario/noologia/</a></p><p>Istituto della Enciclopedia Italiana. (n.d.-c). Polisemìa [Polysemy]. In <em>Vocabolario Treccani</em>. <a href="https://www.treccani.it/vocabolario/polisemia/">https://www.treccani.it/vocabolario/polisemia/</a></p><p>Jaspers, K. (1953). <em>The origin and goal of history</em> (M. Bullock, Trans.). Routledge &amp; Kegan Paul. (Original work published 1949)</p><p>Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., &amp; Fung, P. (2023). Survey of hallucination in natural language generation. <em>ACM Computing Surveys, 55</em>(12), Article 248, 1–38. <a href="https://doi.org/10.1145/3571730">https://doi.org/10.1145/3571730</a></p><p>Jones, C. R., &amp; Bergen, B. K. (2026). Large language models pass a standard three-party Turing test. <em>Proceedings of the National Academy of Sciences, 123</em>(21), e2524472123. <a href="https://doi.org/10.1073/pnas.2524472123">https://doi.org/10.1073/pnas.2524472123</a></p><p>Kleinert, T., Waldschütz, M., Blau, J., Heinrichs, M., &amp; Schiller, B. (2026). AI outperforms humans in establishing interpersonal closeness in emotionally engaging interactions, but only when labelled as human. <em>Communications Psychology, 4</em>, 23. <a href="https://doi.org/10.1038/s44271-025-00391-7">https://doi.org/10.1038/s44271-025-00391-7</a></p><p>Koyré, A. (1957). <em>From the closed world to the infinite universe</em>. Johns Hopkins University Press.</p><p>Laks, A., &amp; Most, G. W. (Eds. &amp; Trans.). (2016a). <em>Early Greek philosophy, Volume III: Early Ionian thinkers, Part 2</em>. Harvard University Press.</p><p>Laks, A., &amp; Most, G. W. (Eds. &amp; Trans.). (2016b). <em>Early Greek philosophy, Volume V: Western Greek thinkers, Part 2</em>. Harvard University Press.</p><p>Laks, A., &amp; Most, G. W. (Eds. &amp; Trans.). (2016c). <em>Early Greek philosophy, Volume VI: Later Ionian and Athenian thinkers, Part 1</em>. Harvard University Press.</p><p>Leibniz, G. W. (1966). <em>Logical papers: A selection</em> (G. H. R. Parkinson, Ed. &amp; Trans.). Clarendon Press.</p><p>Lesher, J. H. (1981). Perceiving and knowing in the <em>Iliad</em> and <em>Odyssey</em>. <em>Phronesis, 26</em>(1), 2–24. <a href="https://doi.org/10.1163/156852881X00105">https://doi.org/10.1163/156852881X00105</a></p><p>Liddell, H. G., Scott, R., Jones, H. S., &amp; McKenzie, R. (1996). <em>A Greek-English lexicon</em> (9th ed., with revised supplement). Clarendon Press.</p><p>Lin, S., Hilton, J., &amp; Evans, O. (2022). TruthfulQA: Measuring how models mimic human falsehoods. In S. Muresan, P. Nakov, &amp; A. Villavicencio (Eds.), <em>Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)</em> (pp. 3214–3252). Association for Computational Linguistics. <a href="https://doi.org/10.18653/v1/2022.acl-long.229">https://doi.org/10.18653/v1/2022.acl-long.229</a></p><p>Liu, J., Li, J., Feng, L., Li, L., Tian, J., &amp; Lee, K. (2014). Seeing Jesus in toast: Neural and behavioral correlates of face pareidolia. <em>Cortex, 53</em>, 60–77. <a href="https://doi.org/10.1016/j.cortex.2014.01.013">https://doi.org/10.1016/j.cortex.2014.01.013</a></p><p>Lloyd, G. E. R., &amp; Sivin, N. (2002). <em>The way and the word: Science and medicine in early China and Greece</em>. Yale University Press.</p><p>Loru, E., Nudo, J., Di Marco, N., Santirocchi, A., Atzeni, R., Cinelli, M., Cestari, V., Rossi-Arnaud, C., &amp; Quattrociocchi, W. (2025). The simulation of judgment in LLMs. <em>Proceedings of the National Academy of Sciences of the United States of America, 122</em>(42), e2518443122. <a href="https://doi.org/10.1073/pnas.2518443122">https://doi.org/10.1073/pnas.2518443122</a></p><p>Manetti, G. (1993). <em>Theories of the sign in classical antiquity</em> (C. Richardson, Trans.). Indiana University Press.</p><p>Matilal, B. K. (1998). <em>The character of logic in India</em>. State University of New York Press.</p><p>McCarthy, J., Minsky, M. L., Rochester, N., &amp; Shannon, C. E. (1955). <em>A proposal for the Dartmouth summer research project on artificial intelligence</em>. Dartmouth College.</p><p>McCulloch, W. S., &amp; Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. <em>Bulletin of Mathematical Biophysics, 5</em>, 115–133. <a href="https://doi.org/10.1007/BF02478259">https://doi.org/10.1007/BF02478259</a></p><p>Nagy, G. (1983). Sēma and Noēsis: Some illustrations. <em>Arethusa, 16</em>, 35–55.</p><p>Nass, C., &amp; Moon, Y. (2000). Machines and mindlessness: Social responses to computers. <em>Journal of Social Issues, 56</em>(1), 81–103. <a href="https://doi.org/10.1111/0022-4537.00153">https://doi.org/10.1111/0022-4537.00153</a></p><p>Nass, C., Steuer, J., &amp; Tauber, E. R. (1994). Computers are social actors. In <em>Proceedings of the SIGCHI Conference on Human Factors in Computing Systems</em> (pp. 72–78). ACM. <a href="https://doi.org/10.1145/191666.191703">https://doi.org/10.1145/191666.191703</a></p><p>Natale, S. (2021). The ELIZA effect: Joseph Weizenbaum and the emergence of chatbots. In <em>Deceitful media: Artificial intelligence and social life after the Turing Test</em> (pp. 50–67). Oxford University Press. <a href="https://doi.org/10.1093/oso/9780190080365.003.0004">https://doi.org/10.1093/oso/9780190080365.003.0004</a></p><p>Needham, J. (1954). <em>Science and civilisation in China. Volume 1: Introductory orientations</em>. Cambridge University Press.</p><p>Needham, J. (1969). <em>The grand titration: Science and society in East and West</em>. George Allen &amp; Unwin.</p><p>Neugebauer, O. (1969). <em>The exact sciences in antiquity</em> (2nd ed.). Dover.</p><p>Plato. (2013a). <em>Republic, Volume I: Books 1–5</em> (C. Emlyn-Jones &amp; W. Preddy, Eds. &amp; Trans.). Harvard University Press.</p><p>Plato. (2013b). <em>Republic, Volume II: Books 6–10</em> (C. Emlyn-Jones &amp; W. Preddy, Eds. &amp; Trans.). Harvard University Press.</p><p>Robson, E. (2008). <em>Mathematics in ancient Iraq: A social history</em>. Princeton University Press.</p><p>Severino, E. (2004). <em>La filosofia dai Greci al nostro tempo. La filosofia contemporanea</em> [Philosophy from the Greeks to our time: Contemporary philosophy]. BUR.</p><p>Severino, E. (2009). <em>Discussioni intorno al senso della verità</em> [Discussions on the meaning of truth]. Edizioni ETS.</p><p>Shannon, C. E. (1948). A mathematical theory of communication. <em>Bell System Technical Journal, 27</em>, 379–423, 623–656. <a href="https://doi.org/10.1002/j.1538-7305.1948.tb01338.x;">https://doi.org/10.1002/j.1538-7305.1948.tb01338.x;</a> <a href="https://doi.org/10.1002/j.1538-7305.1948.tb00917.x">https://doi.org/10.1002/j.1538-7305.1948.tb00917.x</a></p><p>Short, J., Williams, E., &amp; Christie, B. (1976). <em>The social psychology of telecommunications</em>. Wiley.</p><p>Stella, F. (2021). <em>Noos e noein da Omero a Platone</em> [Noos and noein from Homer to Plato]. Presses universitaires de Franche-Comté.</p><p>Turing, A. M. (1936). On computable numbers, with an application to the Entscheidungsproblem. <em>Proceedings of the London Mathematical Society, 42</em>, 230–265. <a href="https://doi.org/10.1112/plms/s2-42.1.230">https://doi.org/10.1112/plms/s2-42.1.230</a></p><p>Turing, A. M. (1950). Computing machinery and intelligence. <em>Mind, 59</em>(236), 433–460. <a href="https://doi.org/10.1093/mind/LIX.236.433">https://doi.org/10.1093/mind/LIX.236.433</a></p><p>Vapnik, V. N. (1995). <em>The nature of statistical learning theory</em>. Springer. <a href="https://doi.org/10.1007/978-1-4757-2440-0">https://doi.org/10.1007/978-1-4757-2440-0</a></p><p>Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., &amp; Polosukhin, I. (2017). Attention is all you need. <em>Advances in Neural Information Processing Systems, 30</em>, 5998–6008.</p><p>Waytz, A., Gray, K., Epley, N., &amp; Wegner, D. M. (2010). Causes and consequences of mind perception. <em>Trends in Cognitive Sciences, 14</em>(8), 383–388. <a href="https://doi.org/10.1016/j.tics.2010.05.006">https://doi.org/10.1016/j.tics.2010.05.006</a></p><p>Weizenbaum, J. (1966). ELIZA: A computer program for the study of natural language communication between man and machine. <em>Communications of the ACM, 9</em>(1), 36–45. <a href="https://doi.org/10.1145/365153.365168">https://doi.org/10.1145/365153.365168</a></p><p>Weizenbaum, J. (1976). <em>Computer power and human reason: From judgment to calculation</em>. W. H. Freeman.</p><p>Wiener, N. (1948). <em>Cybernetics, or control and communication in the animal and the machine</em>. The Technology Press; John Wiley &amp; Sons.</p><p>Wooldridge, M. (2009). <em>An introduction to multiagent systems</em> (2nd ed.). Wiley.</p><p>Xie, S. M., Raghunathan, A., Liang, P., &amp; Ma, T. (2022). An explanation of in-context learning as implicit Bayesian inference. <em>The Tenth International Conference on Learning Representations</em>. <a href="https://openreview.net/forum?id=RdJVFCHjUMI">https://openreview.net/forum?id=RdJVFCHjUMI</a></p><p>Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., &amp; Cao, Y. (2023). ReAct: Synergizing reasoning and acting in language models. <em>The Eleventh International Conference on Learning Representations</em>.</p><p>Zhang, C., Bengio, S., Hardt, M., Recht, B., &amp; Vinyals, O. (2017). Understanding deep learning requires rethinking generalization. <em>International Conference on Learning Representations</em>. <a href="https://openreview.net/forum?id=Sy8gdB9xx">https://openreview.net/forum?id=Sy8gdB9xx</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a4acbb649c39" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When the Machine Seems to Understand]]></title>
            <link>https://medium.com/@enrico.desantis/when-the-machine-seems-to-understand-685ff0de3965?source=rss-e9539446c6d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/685ff0de3965</guid>
            <category><![CDATA[philosophy-of-mind]]></category>
            <category><![CDATA[cognition]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[noosemia]]></category>
            <dc:creator><![CDATA[Enrico De Santis]]></dc:creator>
            <pubDate>Sat, 22 Aug 2026 14:05:40 GMT</pubDate>
            <atom:updated>2026-08-22T14:05:40.762Z</atom:updated>
            <content:encoded><![CDATA[<h3>Semantics, Alignment, and Mind Attribution in Noosemia</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wdu9A7yjwPDRfr5umOOboA.png" /></figure><p><a href="https://medium.com/@enrico.desantis/noosemia-perceiving-the-mind-in-signs-7b23a14c830e"><em>Noosemia</em></a> was introduced to describe a phenomenon that has become increasingly familiar in interactions with advanced generative AI systems. Faced with a machine capable of sustaining a conversation, following a line of reasoning, recognizing implications, producing unexpected connections, and, under certain conditions, developing deductively coherent inferences from given premises (Mondorf &amp; Plank, 2024), users tend to interpret the signs they receive as indications of some form of mental activity. A response may seem to reflect an understanding of context, a suggestion may appear motivated, and an especially pertinent connection may take on the character of an insight. In ordinary language, this shift occurs almost imperceptibly. We say that the system “understood,” “noticed,” “got the point,” or “thought it would be useful” to proceed in a certain direction.</p><p>The concept of <strong>noosemia</strong> — from the Greek <em>noûs</em> (νοῦς), “mind” or “intellect,” and <em>sēmeîon</em> (σημεῖον), “sign,” to denote the perception of a mind through signs — was introduced to delimit this region of experience, in which intentionality, agency, and interiority are progressively attributed to a generative system through the production and interpretation of signs. The ontological question of whether a genuine artificial mind may be present remains a separate question (De Santis &amp; Rizzi, 2025; De Santis, De Santis, &amp; Rizzi, 2025).</p><p>Yet the notion of “projection”, if left unqualified, risks obscuring precisely the most interesting dimension of noosemia. The term may suggest a human subject freely imposing mental states on a computational object, as though the attribution depended primarily on the observer’s psychological disposition. Such an account captures a genuine component, documented by a long tradition of research on anthropomorphism, but becomes insufficient when the object of interaction is an extremely complex system capable of producing language that is highly sensitive to context, adapting to the interlocutor’s corrections, and generating conceptual relations that the user recognizes as relevant even when they had not previously formulated them.</p><p>The analysis must therefore move beneath the phenomenological surface of conversation. Fluency, dialogical coherence, first-person language, and other anthropomorphic cues can measurably affect how a system is perceived. In a large experimental study using a pseudo-LLM, for example, a speech-plus-text interface increased anthropomorphism and perceived information accuracy, while first-person framing produced more context-dependent effects on accuracy and risk judgments (Cohn et al., 2024). These features belong to the surface of interaction. A deeper component of noosemia seems to emerge when users perceive the machine as operating within a domain of meanings sufficiently close to their own to make the impression of understanding plausible. What requires more careful explanation is the transition from the relevance of a sign to the perception that someone has understood.</p><h3>The Sign and the Question of Who Understands</h3><p>The question has a genealogy far older than contemporary generative systems. In Charles Sanders Peirce’s semiotics, a sign belongs to a relation in which something stands for something else in some respect and produces an interpretant, that is, a further semiotic effect through which that relation is developed and may enter into a new interpretation (Peirce, 1998; Short, 2007).</p><p>The Peircean interpretant should be understood as the effect through which a sign acquires further determination within a potentially open-ended process of semiosis, in which an interpretant may itself become a further sign. Umberto Eco later developed this Peircean line of thought through the notion of <em>unlimited semiosis</em>, understood as the potentially open-ended character of interpretation rather than as the absence of interpretive constraints (Eco, 1992).</p><p>This perspective is especially fruitful for the analysis of generative systems because it allows meaning to be considered as something that circulates and transforms through signs while leaving temporarily suspended the question of whether a conscious subject must be posited as the origin of that process. The latter question is itself notoriously difficult, given the absence of rigorous and widely shared definitions and, consequently, of a scientifically established empirical framework.</p><p>In a conversation with a large language model (LLM), the sign produced by the user is processed through the system’s representational organization and gives rise to a new sign. This new sign elicits further interpretants in the user, which may in turn be expressed linguistically and returned to the machine on the next turn. Interaction thus takes the form of a <strong>semiotic recursion</strong>, in which interpretive effects arising at one pole of the relation become conditions for subsequent sign production.</p><p>Such a dynamic can unfold without first determining whether the two poles possess the same mental status. This suspension is precisely what makes semiosis useful for understanding noosemia. Over the course of the recursion, however, the user may begin to experience the machine as a source of interpretations rather than merely as a medium through which signs are transformed. The distance between these two readings is subtle. In the first, one observes a semantically effective transformation; in the second, something that interprets is implicitly posited. Noosemia takes shape in the space between these two levels.</p><h3>Two Heterogeneous Semantic Spaces</h3><p>To approach the problem, it may be useful, with due caution, to invoke the metaphor of two semantic networks.</p><p>Human conceptual organization is shaped by a singular history of experience. Language and perception are intertwined with embodiment and autobiographical memory, while social learning and knowledge acquired throughout life contribute to an ever-changing web of relations among concepts. Some associations are deeply consolidated, whereas others remain weak or latent. Certain concepts are closely connected, while others may belong to the same individual experience without ever having been explicitly related. The role of perception, situated action, and bodily states in the organization of cognition has been extensively discussed within the framework of <strong>grounded cognition</strong> (Barsalou, 2008).</p><p>Cognitive science has long attempted to model aspects of this relational organization. Collins and Loftus’s spreading-activation theory described semantic memory through structures in which activation of one concept could propagate toward semantically connected concepts (Collins &amp; Loftus, 1975). Later models, such as Latent Semantic Analysis, showed how semantically rich organizations could emerge statistically from regularities in large textual corpora, opening a line of research that would contribute to modern distributional semantics (Landauer &amp; Dumais, 1997; Boleda, 2020) and to techniques for representing linguistic elements as embeddings in structured mathematical spaces used in machine learning.</p><p>Contemporary large language models likewise develop highly articulated representational structures. During training, vast amounts of linguistic regularity are incorporated into the parameters of the neural network; in multimodal systems, such regularities may also derive from images, sounds, and other modalities. This process gives rise to representational configurations capable of relating words, objects, properties, events, and abstract structures or schemas, making dependencies available for use during generation.</p><p>Identifying this organization with human semantics would be unwarranted. The two have different origins and belong to profoundly heterogeneous systems. Human meaning develops within an embodied, perceptual, and social existence; model representations emerge from optimization processes through which the system learns regularities and long-range correlations in data. This difference is precisely what keeps the problem of grounding open.</p><p>Harnad formulated the <strong>symbol grounding problem</strong> by asking how the semantic interpretation of a symbolic system could become intrinsic to the system itself rather than depending on meanings already present in the mind of an interpreter (Harnad, 1990). Many years later, Bender and Koller developed a related distinction, arguing that learning linguistic form alone does not, in itself, provide access to the relation among expression, communicative intent, and the world that characterizes their conception of meaning (Bender &amp; Koller, 2020).</p><p>Contemporary language models nevertheless make it difficult to frame the issue as a stark opposition. Their representational organizations exhibit extremely rich conceptual relations, and several studies have identified measurable correspondences between certain artificial structures and those observed in humans. Goldstein and colleagues found common geometric patterns between contextual embeddings from a deep language model and neural representations sampled from the inferior frontal gyrus. Mischler and colleagues reported converging aspects of hierarchical feature extraction in language models and human neural language processing. In a different domain, Kawakita and collaborators used Gromov–Wasserstein optimal transport to compare human and artificial color-similarity structures, identifying structural correspondences without assuming that the two spaces shared the same representational coordinates (Goldstein et al., 2024; Kawakita et al., 2024; Mischler et al., 2024).</p><p>The notion of structural similarity nevertheless requires particular caution. A correspondence between representational geometries indicates that certain internal relations may be organized according to comparable configurations, but it does not imply that the related elements have the same semantic status, the same causal origin, or the same form of access to the world. Two systems may display partially corresponding relational organizations while remaining radically different in how those relations are constituted and used.</p><p>Within the framework proposed here, <strong>structural similarity</strong> should be distinguished from <strong>semantic interoperability</strong>. Structural similarity concerns the possibility of identifying correspondences between relational organizations; semantic interoperability arises when those correspondences become operationally productive within an interaction. Two representational spaces may be considered locally interoperable when a relation expressed in one is transformed into a sign that, once processed by the other system, preserves enough structure to produce a relevant response, a new inference, or a coherent modification of subsequent behavior.</p><p>Interoperability thus describes a partial functional compatibility through which heterogeneous systems can coordinate portions of their relational spaces. A semiotic perspective allows this process to be understood without locating the entirety of meaningfulness in either system, since part of what the user experiences as meaning emerges along the chain of transformations through which a sign generates further interpretants, re-enters the conversation, and modifies what can be produced in subsequent turns. Within this local compatibility, continually reactivated through dialogical semiosis, noosemia finds one of its deeper conditions.</p><h3>Resonance, Difference, and Surprise</h3><p>When the user formulates a question, they do not provide the system with a neutral sequence of words. They introduce a fragment of their own conceptual space. The vocabulary they choose, the distinctions they draw, the examples they provide, and what they leave implicit delimit a field of relations that the model uses to construct its response.</p><p>A first form of resonance emerges when artificial output preserves relations already present in the user’s conceptual space. The machine associates concepts that the user regards as naturally connected, develops inferences recognized as plausible, and preserves distinctions regarded as semantically relevant.</p><p>Under certain conditions, this competence also includes inferences whose structure respects deductive relations. This qualification matters because the inductive character primarily concerns the learning process, during which regularities are statistically acquired from data. Once acquired, however, these structures can support outputs that correctly instantiate relations characteristic of deductive logic.</p><p>In propositional-logic tasks, experimental studies show that the language models examined can employ inferential strategies associated with deductive reasoning, including patterns of supposition following and chain construction. At the same time, a correct final answer does not by itself establish that the reasoning process producing it was logically sound, stable, or generalizable (Mondorf &amp; Plank, 2024). For noosemia, however, what matters in the interlocutor’s immediate experience is also the recognizable structure of the inference. When the machine derives a logically coherent consequence from premises supplied by the user, especially when that consequence was not explicitly stated, the output may be perceived as the manifestation of an understanding capable of developing the received content autonomously.</p><p>Resonance involves more than simple overlap. If the interaction merely returned to users what they already thought, it would be unlikely to produce some of the strongest forms of mind attribution. The phenomenon becomes more interesting when a <strong>productive difference</strong> emerges.</p><p>A user may possess two concepts without ever having established a salient relation between them. By virtue of its training history and representational organization, the model may instead produce a connection between those same elements. The user evaluates it and finds the relation relevant. It may take the form of an analogy, a theoretical consequence, a historical precedent, or an interdisciplinary connection that had not previously been considered. In modern <strong>agentic AI systems</strong>, this may also occur after the completion of a complex task, such as producing working code that satisfies the user’s requirements through an automated sequence of generation, goal refinement, and autonomous testing.</p><p>Surprise has a distinctive character in this setting. A wholly alien connection would have little significance; what makes the output surprising is the conjunction of novelty and recognizability. The system produces something the user had not thought of, yet once formulated it appears compatible with the user’s conceptual world.</p><p>This dynamic can be described as <strong>productive relational surprise</strong>. It arises when a relation that was initially of low salience within the user’s conceptual space is proposed by the model and subsequently recognized as coherent, fruitful, or explanatory. Novelty stems from the difference between the two semantic organizations, whereas meaningfulness depends on their partial interoperability. It is this combination that gives the experience its force.</p><p>The machine seems to have “seen” a relation, and the anthropomorphic verb already signals an interpretive shift: a property of sign production is being described through a category ordinarily applied to a cognitive subject. The user might simply note that a particular relation is semantically well-formed; in lived interaction, however, the impression may arise that the system has grasped something the user had missed.</p><p>Productive difference thereby acquires a distinctive role. Near-perfect correspondence between the two structures would reduce the model’s capacity to surprise, whereas excessive distance would make its outputs irrelevant or incomprehensible. The most fertile regime for noosemic experience may therefore lie between these extremes, where the artificial structure is similar enough to remain interpretable yet different enough to introduce unexpected relations.</p><p>Error, too, has a phenomenologically ambiguous role within this framework. An obvious hallucination, loss of context, or manifestly incongruous response can interrupt the continuity of interaction and foster the inverse dynamic termed <strong>a-noosemia</strong>: the weakening or collapse of mind attribution, as the presence previously perceived through the signs recedes and the system once again appears primarily as a device or tool (De Santis &amp; Rizzi, 2025).</p><p>Within the interpretive framework proposed here, however, error can also have the opposite effect. When it takes a form that the user recognizes as structurally similar to human fallibility — an intuitively plausible but incorrect inference, a systematic bias in judgment, an unwarranted generalization, or a contextually plausible lapse — it may be interpreted as the error of a subject rather than merely as a malfunction of a mechanism. Studies of syllogistic reasoning have identified systematic errors and reasoning biases that, at the behavioral level, display similarities to those observed in humans (Ozeki et al., 2024). Such correspondences do not warrant inferences about the identity of the underlying cognitive processes and, considered in isolation, do not establish that fallibility increases mind attribution. They nevertheless suggest an interesting condition for noosemia, since even an error may take a form recognizable within familiar categories of human fallibility.</p><p>Noosemia may thus intensify within a delicate region of balance in which the machine is similar enough to be understood, different enough to surprise, and, under certain conditions, even fallible in ways the interlocutor recognizes as familiar.</p><h3>Dialogue as a Process of Alignment</h3><p>Prolonged interaction introduces a further transformation. The two semantic organizations do not remain unchanged as the conversation unfolds.</p><p>The user corrects the system, specifies the intended meaning of a term, establishes analogies, rules out unwanted interpretations, and introduces new relations. They may explain that one notion should be treated as subordinate to another, that two apparently similar terms must remain distinct, or that a concept used in the conversation carries a technical meaning different from its ordinary one. <strong>Contextual steering</strong> continuously introduces structure into the communicative flow.</p><p>In ordinary conversational use, context changes the conditions under which generation occurs without requiring online updates to the neural model’s parameters. The representations activated on the next turn depend on the previous history of the interaction and can therefore be progressively steered toward regions of representational space that are more compatible with the conceptual framework introduced by the user (Shanahan, 2024).</p><p>A complementary process occurs on the human side. A relation proposed by the model may be accepted, rejected, or modified. Once accepted, it may enter the organization through which the user interprets subsequent turns. Conversation therefore changes both systems, but in profoundly different ways. The language model changes its local contextual state, whereas the human interlocutor may alter their conceptual organization more persistently.</p><p>Research on human dialogue provides a useful point of comparison. Pickering and Garrod described human dialogue through the notion of <strong>interactive alignment</strong>, according to which the representations used by interlocutors tend to align at different levels during interaction (Pickering &amp; Garrod, 2004). Conversation also constructs <strong>common ground</strong>, which is continually established and updated so that participants can coordinate meaning effectively (Clark &amp; Brennan, 1991).</p><p>Interaction with LLMs presents a distinctive configuration of this process. In the dialogue settings they examined, Shaikh and colleagues found that LLM generations contained substantially fewer explicit grounding acts than human utterances and often behaved as though common ground were already available (Shaikh et al., 2024). In a complementary line of work, Cazalets and collaborators found that humans adapted their communicative strategies when interacting with an LLM and that this adaptation occurred irrespective of whether participants were aware that their interlocutor was artificial (Cazalets et al., 2025).</p><p>These findings suggest a bidirectional dynamic that requires careful description. The user steers the model through context while simultaneously adapting their own behavior to the responses received. Over many turns, <strong>a local, jointly constituted semantic field</strong> may thereby emerge in which certain distinctions, assumptions, and relations are <strong>operationally treated as shared</strong>. Global convergence between the two structures is unnecessary. It is more appropriate to speak of <strong>local and contingent semantic alignment</strong>, because the conversation selects a limited region of the respective conceptual spaces and reduces, within that region, certain functionally relevant differences. Interaction becomes more effective insofar as each new turn preserves previously established relations and compositionally builds on them in subsequent turns.</p><p>From a semiotic perspective, this dynamic can also be described as a progressive stabilization of interpretants. Some distinctions introduced by the user are taken up by the system, reappear in transformed form in later outputs, and return to the user as signs that confirm, modify, or extend the previous interpretive framework. The meaning of the conversation thus resides, in part, in the history of transformations made possible by each sign.</p><p>When the system correctly uses a distinction introduced several turns earlier, extends a line of reasoning while respecting constraints established by the user, or draws out an implication that had not been explicitly stated, the experience tends to be interpreted through the ordinary vocabulary of understanding. The expression “it understood what I meant” marks precisely the transition from a relation between structures to a relation between perceived subjects.</p><h3>Perceived Understanding</h3><p>To say that a response is relevant describes a relation among a request, a context, and an output. To say that someone has understood introduces a subject to whom an epistemic state is attributed.</p><p>Between these two descriptions lies an inferential space that, in everyday human interaction, we traverse with little difficulty. If a person responds appropriately to what we say, preserves the implications of the discussion, integrates our corrections, and modifies their behavior in light of the clarifications they receive, we normally assume that they have understood. In interaction with an LLM, the same cognitive heuristic is applied to an entity of a different kind.</p><p>A distinction must therefore be drawn between <strong>actual understanding</strong>, however one chooses to define it philosophically, and <strong>perceived understanding</strong>. The latter designates the degree to which the user interprets the system’s behavior as manifesting appropriate access to the meaning of their discourse. Noosemia primarily concerns this second dimension, which tends to strengthen when the system displays stable sensitivity to the relations structuring the user’s discourse. The correctness of a response contributes to this impression, which becomes stronger when combined with temporal continuity, the preservation of local distinctions, the production of relevant inferences, and responsiveness to correction.</p><p>Anthropomorphism provides part of the explanation. Epley, Waytz, and Cacioppo showed that the attribution of human properties to nonhuman agents depends, among other factors, on the availability of anthropocentric knowledge through which behavior can be interpreted and on the need to make that behavior understandable and predictable (Epley et al., 2007). Mind perception also comprises differentiated dimensions. Gray, Gray, and Wegner experimentally distinguished two dimensions of mind perception, <strong>Agency</strong> and <strong>Experience</strong>, showing that an entity may be regarded as capable of planning or self-control without being credited to the same extent with experiential capacities such as feeling pain or pleasure (Gray et al., 2007).</p><p>For noosemia, this distinction matters because users may perceive cognitive competence, understanding, and agency in a model without forming any belief about the existence of phenomenal consciousness. Mind attribution therefore appears graded and multidimensional.</p><p>Perceived understanding plays a mediating role. Structural similarity and dialogical alignment, considered in isolation, do not warrant mind attribution; they become phenomenologically relevant when interpreted as evidence that the system has access to the meaning organizing the conversation. The implicit inference can be stated in elementary form. “The response is compatible with what I meant” becomes “the system understood what I meant.” Through this shift, observed appropriateness is traced back to a hypothesized internal state, preparing the transition toward the attribution of agency.</p><h3>From Understanding to Agency</h3><p>Understanding and agency should remain analytically distinct. A system may be perceived as capable of understanding a request while still appearing entirely reactive, merely responding within the boundaries defined by the user. The impression of agency emerges when behavior begins to be interpreted in terms of selection among possibilities, maintenance of a goal, initiative, or goal-directed adaptation. Perceived understanding makes this second transition possible because it provides the epistemic precondition for interpreting behavior as a relevant choice.</p><p>If the system spontaneously suggests a line of further inquiry before the user requests it, the output can be described simply as a plausible continuation of the conversation. Once the user already believes that the system has understood the problem, however, the same behavior takes on a different significance and the suggestion may appear to result from an evaluation. “It understood the problem” can thus become “it judged this worth suggesting.”</p><p>The shift is modest in language but substantial at the cognitive level. Perceived understanding attributes access to content to the system, whereas perceived agency introduces the idea that this content may guide its subsequent behavior.</p><p>Here the phenomenology of intentionality offers a useful distinction. In the Husserlian tradition, intentionality designates the character by which consciousness is always consciousness <em>of</em> something, that is, its constitutive <strong>directedness</strong> toward an object, a content, or a state of affairs (Husserl, 1913/1982). The concept therefore concerns a directedness proper to experience and has a more specific meaning than the ordinary notion of intention as a deliberate plan.</p><p>This distinction helps clarify the noosemic dimension. The user begins to perceive the machine’s actions as directed toward something. A response appears to address the problem they are trying to solve; a suggestion seems oriented toward a possible development of the conversation; a request for clarification appears directed toward a recognized informational gap.</p><p>Behavior thereby acquires a phenomenologically legible directedness. Whether intentionality in the Husserlian sense is actually present remains a separate question. What matters here is that the system’s conduct can be experienced through an interpretive form that has the structure of intentionality. The output appears to be about something, while its unfolding appears oriented by what it is about.</p><p>The noosemic progression can therefore be described as a sequence in which semantic congruence makes perceived understanding possible, perceived understanding prepares the attribution of agency, and agency opens the way to an intentional interpretation of behavior. At this level, a correction may appear to result from the system having “noticed” a problem; a request for clarification may seem motivated by an awareness that information is missing; and a suggestion may be interpreted as a choice oriented toward the interlocutor’s benefit. The system thus appears capable of acting on the basis of what it has understood.</p><h3>When Behavior Comes to Seem Intentional</h3><p>Daniel Dennett’s <strong>intentional stance</strong> provides an especially suitable framework for understanding this transformation. According to Dennett, a system may be interpreted by attributing beliefs, desires, and intentions to it when this strategy makes its behavior effectively predictable (Dennett, 1987).</p><p>Dennett’s approach and Husserl’s phenomenological framework belong to different traditions and should remain distinct. The former concerns an interpretive strategy through which an observer renders a system’s behavior intelligible; the latter describes the intentional structure of experience. Their juxtaposition is nevertheless illuminating in the case of noosemia. The user progressively adopts an intentional interpretation of the system when artificial behavior appears organized as though directed toward objects, problems, goals, and shared meanings. This interpretive tendency may become especially salient in advanced systems capable of <strong>tool use</strong> and the autonomous execution of complex tasks.</p><p>In interaction with an LLM, the intentional stance can be particularly effective. Saying that the model “understood that a piece of information was missing” may be cognitively more manageable than describing the combination of contextual representations and internal transformations that produced a request for clarification. Similarly, saying that it “decided to explore that point further” offers a compact account of a sequence that, at the computational level, results from a generative process conditioned by multiple factors. Mentalistic language thereby provides an efficient explanatory shorthand.</p><p>Shanahan has emphasized the need to keep this functional vocabulary distinct from an ontological reading of it. Describing a language model with words such as “know,” “believe,” or “want” can be useful, provided that such usage is not transformed, without further argument, into the claim that the system possesses those states in the same sense in which we attribute them to humans. The notion of <strong>role play</strong> proposed by Shanahan, McDonell, and Reynolds allows for a high-level description of model behavior while avoiding an automatic transition from observable performance to literal psychology (Shanahan et al., 2023; Shanahan, 2024).</p><p>Noosemia operates at a different level: its primary object is the user’s experience when this vocabulary begins to feel spontaneously appropriate. The system produces signs organized according to relations the user recognizes; conversation locally increases compatibility between the active structures of the two interlocutors; the model sometimes generates unexpected relations that the user judges valid; and dialogical continuity consolidates the impression that common ground exists. When congruence is interpreted as understanding, and understanding as a principle capable of guiding behavior, the mentalistic explanation progressively acquires phenomenological plausibility.</p><p>A mind is then perceived through the signs because those signs appear organized as though they came from something that understands and acts on the basis of what it has understood.</p><h3>The Explanatory Gap</h3><p>A further element, already present in the original formulation of noosemia, amplifies this dynamic (De Santis &amp; Rizzi, 2025). The user is confronted with the outcome of processes that remain largely inaccessible. They may understand in general terms how a <a href="https://medium.com/@enrico.desantis/why-transformers-changed-language-modeling-76d6fee8e398">Transformer architecture</a> works, grasp the principle of attention, or know that generation depends on distributions conditioned on prior context. Even a technically sophisticated user, however, is unlikely to reconstruct in full why a particular internal configuration gave rise to precisely a certain analogy, suggestion, or inference at that particular moment in the conversation.</p><p>This discrepancy can be described through the notion of an <strong>explanatory gap</strong>, understood here as the distance between the intelligibility of observable behavior and the accessibility of the causal mechanism that generates it (De Santis &amp; Rizzi, 2025; De Santis, De Santis, &amp; Rizzi, 2025). In this context, the expression has an <strong>epistemic and mechanistic</strong> meaning and must be distinguished from the <em>explanatory gap</em> discussed in the philosophy of consciousness. The issue here concerns the difficulty of moving from the system’s manifest behavior to an intelligible reconstruction of the dynamics that produced it.</p><p>A notable antecedent to this distinction <a href="https://medium.com/@enrico.desantis/from-the-turing-test-to-noosemia-when-the-sign-seems-to-have-a-mind-6cf27563c74e">can already be found in Alan Turing</a>. In 1950, while discussing Lady Lovelace’s objection that a machine could not produce anything genuinely new, Turing directly addressed the problem of surprise. He observed that machines surprised him with great frequency and attributed this surprise to the fact that he had not performed enough calculation to anticipate their behavior (Turing, 1950, pp. 450–451).</p><p>Read in the present context, Turing’s observation suggests that machine-generated surprise can arise from a disproportion between what the system can process and what a human observer can practically reconstruct, even within a fully computational causal framework.</p><p>Two years later, during the BBC radio discussion <em>Can Automatic Calculating Machines Be Said to Think?</em>, Turing made the point still more precise. Geoffrey Jefferson had suggested that the difficulty arose from our lack of knowledge of the “wiring” of the human mind. Turing rejected the analogy, noting that in the case of the machine the wiring was known and yet unexpected behavior could still occur. In principle, such behavior could have been predicted; in practice, Turing explained, making that prediction would generally require too much work. He added that if one had to anticipate every operation of a computer by reproducing its entire process in advance, part of the reason for using a computer in the first place would disappear (Turing et al., 1952/2004, p. 500).</p><p>The distinction anticipated by Turing is epistemologically important. Knowing the rules governing a system is not equivalent to being able to predict all of its behavior with ease. Causal determination and predictability in practice belong to different levels. A process can be entirely produced by a known computational structure and yet remain too costly to retrace for an observer to maintain an intuitively accessible causal explanation.</p><p>In contemporary generative systems, this distance takes on a more complex form. To the quantitative disproportion already glimpsed by Turing are added the distributed organization of representations, the nonlinearity of transformations, and the contextual dependence of internal processes. Knowing the architecture, training procedures, and even the parameters of a model does not amount to possessing a transparent explanation of each particular behavior.</p><p>This creates a striking asymmetry. The output may be immediately intelligible at the semantic level, while the dynamics that generated it remain far less accessible at the causal level. We understand what the machine is saying much more readily than we can reconstruct why it produced precisely that response, that connection, or that particular inferential trajectory.</p><p>The gap has at least two closely related dimensions. The first concerns <strong>how</strong>, because the user normally lacks a transparent reconstruction of the causal path through which the system arrived at a given output. The second concerns <strong>why</strong>, because relevant and apparently directed behavior invites an account capable of rendering that direction intelligible.</p><p>It is particularly within this second dimension that intentional vocabulary offers a powerful form of <strong>explanatory compression</strong>. To say that the system “understood the problem and judged it appropriate to suggest that solution” organizes, through cognitive categories, a computational dynamic that is much more difficult to reconstruct in its actual steps.</p><p>The explanatory gap does not, by itself, produce noosemia. The opacity of a process that generated semantically incoherent results would be unlikely to suggest the presence of a mind. Its role is instead to amplify an attribution already made plausible by semantic resonance, productive surprise, and continuity of interaction.</p><p>The account proposed here suggests that, as the distance between the intelligibility of the outcome and the transparency of the process increases, an intentional description may become a cognitively economical way of rendering the observed behavior intelligible. On this account, complexity enters the noosemic process indirectly, increasing the distance between what the system makes manifest and what the observer can causally reconstruct, and leaving open an interpretive space in which categories such as understanding, choice, insight, and intention become more readily available.</p><h3>A Possible Form of Functional Subjectivity</h3><p>Up to this point, the process can be described by focusing on the relation between machine behavior and user interpretation. The contemporary picture becomes more complex if some of the properties users interpret as signs of a subject correspond to forms of functional organization actually present in the system.</p><p>This possibility calls for a distinction between <a href="https://medium.com/@enrico.desantis/functional-subjectivity-as-an-operative-constraint-autorecursivity-language-and-memory-in-ae6495a20aea"><strong>functional subjectivity</strong></a> and <strong>phenomenal subjectivity</strong>. Functional subjectivity can be understood as an operational configuration in which different processes are coordinated around a relative continuity of perspective. Such continuity may be expressed in the stability of the conversational persona, functional access to certain internal representations, the capacity to maintain self-referential constraints, or forms of self-monitoring that causally affect behavior (De Santis, 2026).</p><p>Phenomenal subjectivity, by contrast, concerns genuine subjective experience. Identifying a particular functional organization leaves open whether the system undergoes any subjective experience at all. The relation between these two levels therefore remains unresolved, and the presence of functional properties associated with self-reference, continuity, or internal access does not settle the ontological question of consciousness.</p><p>Recent interpretability research nevertheless makes this distinction empirically salient. Lindsey (2025) reported <strong>limited, unreliable, and context-dependent evidence of functional introspective awareness</strong> in the models tested. Through controlled interventions on internal activations, the experiments showed that some models could, under specific conditions, detect and identify injected representations and exhibit behavior consistent with a degree of control over internal states. These findings concern a circumscribed functional capacity and do not establish phenomenal consciousness.</p><p>A complementary line of research concerns the internal organization of persona. Lu and colleagues identified an <strong>Assistant Axis</strong>, the leading direction in a representational persona space across several models. Causal steering along this axis altered behavior associated with the default Assistant persona, while deviations from it predicted <strong>persona drift</strong> (Lu et al., 2026). These results provide evidence that aspects of conversational persona have internal representational correlates capable of causally influencing behavior.</p><p>More recently, Gurnee and collaborators identified a privileged space of verbalizable representations, termed <strong>J-space</strong>, that exhibits several <strong>functional properties characteristic of a global workspace</strong>. Its contents are available for verbal report, can support intermediate stages of reasoning, and can be made available to downstream computations. The authors further report that post-training installs what they call <strong>the Assistant’s point of view</strong> in this workspace (Gurnee et al., 2026).</p><p>These findings call for caution in both directions. Moving from functional correlates of self-reference, continuity, or internal access to a claim about consciousness remains unjustified. At the same time, it would be equally unwarranted to assume that every property perceived by the user as mentalistic must lack any <strong>internal functional correlate</strong>.</p><p>Functional subjectivity enters the argument at precisely this level. It constitutes a <strong>possible causal source</strong> of some of the properties through which the noosemic effect manifests itself, while leaving the phenomenological center of gravity of noosemia unchanged. A more stable conversational configuration can strengthen temporal continuity; forms of functional access to internal representations can increase the coherence of self-reference; monitoring mechanisms can produce corrections or adjustments that, from the outside, are interpreted as manifestations of reflexivity.</p><p>The user does not directly perceive these structures but encounters their effects on behavior. The relation between functional subjectivity and noosemia should therefore be understood as indirect and causal. An internal organization capable of sustaining greater stability and coherence may strengthen perceived understanding and agency and, through them, make intentional attribution more persistent. The possibility of functional subjectivity thus makes the notion of projection more complex while leaving the ontological question of artificial mind unresolved.</p><h3>Attribution Under Constraint</h3><p>In light of this reconstruction, the term “projection” requires greater precision.</p><p>Through their prompts, users introduce into the interaction categories drawn from their own experience of mind. They have no direct access to any possible artificial interiority and interpret behavior through cognitive schemas developed primarily in relations with other human beings. Yet this attribution takes place under real constraints.</p><p>The model preserves semantic relations, responds to corrections, builds contextual continuity, and sometimes produces unexpected connections that the user recognizes as valid. During dialogue, both systems contribute to reshaping the field in which the conversation unfolds. Perceived common ground may increase, while the operational distance between the semantically active structures may locally decrease.</p><p>At the same time, the<strong> complexity of the system</strong> makes it difficult to reconstruct the causal process leading to each behavior. In some advanced models, internal properties functionally associated with the continuity of an artificial persona or with limited forms of self-monitoring may also contribute to the stability of the interaction.</p><p>Noosemia emerges from the intersection of these levels without coinciding with any one of them. Users may be fully aware of the computational nature of the system and, at the same time, experience the conversation as an encounter with something that understands, anticipates, remembers, and orients its behavior in meaningful ways. This tension between reflective knowledge and immediate experience is one of the most distinctive features of the phenomenon.</p><p>The mechanistic description remains available in the background, while the intentional description becomes increasingly natural and parsimonious as an experiential description. For this reason, it appears more appropriate to speak of <strong>constrained interpretive attribution</strong> than of simple projection. The term “projection” can still be used, but only with the qualifications developed here. Attribution arises from the encounter between the observer’s cognitive schemas and actual properties of behavior, while the internal causal structure remains only partially accessible. Noosemia thus belongs to the relation between evidence and interpretation.</p><h3>From Sign to Presence</h3><p>We can now specify more precisely what happens when we say that a generative system “seems to have a mind.”</p><p>The surface of dialogue provides the first cues. Language is fluent, the response is relevant, and the form of interaction resembles human conversation. Beneath this surface operates a deeper dynamic in which two heterogeneous semantic organizations enter into relation. Some configurations prove structurally compatible and, when this compatibility becomes productive in dialogue, a local form of semantic interoperability emerges. The model returns relations that the user recognizes and introduces others that may be surprising without appearing arbitrary. Contextual steering progressively reduces some local differences, while the history of the conversation builds an increasingly structured shared field.</p><p>The semiotic structure of interaction adds another level. Every artificial output becomes a sign for the user, generates interpretants, and may eventually return to the machine as new context. Over the course of this recursion, the user’s interpretive history becomes progressively incorporated into the conditions influencing subsequent outputs, so that the system returns signs that reflect, at least in part, interpretations previously introduced by the interlocutor.</p><p>From this circularity, a particularly subtle phenomenological transformation may emerge. What initially appeared as the transformation of a sign may come to be experienced as interpretation; contextual continuity may be experienced as memory; sensitivity to relations may take the form of understanding. When such understanding seems to orient behavior toward something, the structure of attributed intentionality emerges.</p><p>Semantic congruence is thus interpreted as understanding; perceived understanding makes the attribution of agency possible, because behavior consistent with a meaning already believed to have been understood can be read as guided by that meaning; agency, in turn, opens onto an intentional interpretation in which what the system does appears explicable through hypothesized internal states or processes — understanding, remembering, evaluating, choosing.</p><p>The <strong>explanatory gap</strong> further amplifies this movement. Turing had already identified a distance between knowledge of the machine and the practical possibility of anticipating all of its operations; in contemporary generative systems, that distance is increased by representational complexity and contextual dependence. Error also contributes ambivalently to the dynamic: it may interrupt attribution when it breaks the coherence of interaction, or it may assume a form of fallibility that the user recognizes as familiar. The possible presence of structures of functional subjectivity adds a further qualification, since some behaviors that sustain attribution may be supported by internal organizations oriented toward continuity, self-monitoring, or the stabilization of an operational perspective, without any conclusion about phenomenal experience following from this.</p><p>Noosemia can therefore be described as the process through which <strong>the dynamic interoperability of heterogeneous semantic structures, sustained by progressively stabilized semiotic recursion, is interpreted and experienced by the human user as intentional reciprocity</strong>.</p><p>This formulation makes it possible to distinguish noosemia both from anthropomorphism grounded primarily in aesthetic or behavioral resemblance and from the ontological claim that the machine necessarily possesses a mind. The object of noosemia remains the interpretive interface between the human interlocutor and the artificial system: the locus at which actual properties of behavior, possible internal functional organizations, and the observer’s cognitive categories meet to the point of destabilizing the boundary between sign and subject.</p><p>The semiotic perspective helps explain how meaning can develop through a chain of transformations; phenomenology clarifies why, at a certain point, that chain may be experienced as originating in something that appears directed toward the world; the <strong>intentional stance</strong>, in turn, explains why a mentalistic description can become an especially effective interpretive strategy. The specificity of noosemia lies at the intersection of these levels.</p><p>The question, then, is how an artificial structure capable of operating within a space of relations semantically interoperable with our own can progressively become, in experience, the locus from which meaning seems to arise in an intentional form perceived as presence. The transition from “this response is meaningful” to “someone has understood what it means” condenses a minimal shift in language and a far deeper one in philosophical terms.</p><p><strong>Noosemia lies within this interpretive shift.</strong> As technology advances, the same shift may open new horizons of meaning in which a new form of perceived mind becomes intelligible as genuine presence.</p><h3>References</h3><p>Barsalou, L. W. (2008). Grounded cognition. Annual Review of Psychology, 59, 617–645. doi 10.1146/annurev.psych.59.103006.093639</p><p>Bender, E. M., &amp; Koller, A. (2020). Climbing towards NLU: On meaning, form, and understanding in the age of data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 5185–5198). Association for Computational Linguistics. doi 10.18653/v1/2020.acl-main.463</p><p>Boleda, G. (2020). Distributional semantics and linguistic theory. Annual Review of Linguistics, 6, 213–234. doi 10.1146/annurev-linguistics-011619–030303</p><p>Cazalets, T., Janssens, R., Belpaeme, T., &amp; Dambre, J. (2025). Human alignment: How much do we adapt to LLMs? In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) (pp. 603–613). Association for Computational Linguistics. doi 10.18653/v1/2025.acl-short.47</p><p>Clark, H. H., &amp; Brennan, S. E. (1991). Grounding in communication. In L. B. Resnick, J. M. Levine, &amp; S. D. Teasley (Eds.), Perspectives on socially shared cognition (pp. 127–149). American Psychological Association. doi 10.1037/10096–006</p><p>Cohn, M., Pushkarna, M., Olanubi, G. O., Moran, J. M., Padgett, D., Mengesha, Z., &amp; Heldreth, C. (2024). Believing anthropomorphism: Examining the role of anthropomorphic cues on trust in large language models. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems (Article 54, pp. 1–15). Association for Computing Machinery. doi 10.1145/3613905.3650818</p><p>Collins, A. M., &amp; Loftus, E. F. (1975). A spreading-activation theory of semantic processing. Psychological Review, 82(6), 407–428. doi 10.1037/0033–295X.82.6.407</p><p>De Santis, E. (2026, January 7). Functional subjectivity as an operative constraint: Autorecursivity, language, and memory in advanced language models. Medium.</p><p>De Santis, E., De Santis, G., &amp; Rizzi, A. (2025). Noosemia. Prometeo, 43(172), 16–24.</p><p>De Santis, E., &amp; Rizzi, A. (2025). Noosemia: Toward a cognitive and phenomenological account of intentionality attribution in human–generative AI interaction. arXiv. arXiv:2508.02622. doi 10.48550/arXiv.2508.02622</p><p>Dennett, D. C. (1987). The intentional stance. MIT Press.</p><p>Eco, U. (1992). Interpretation and history. In S. Collini (Ed.), Interpretation and overinterpretation (pp. 23–44). Cambridge University Press. doi 10.1017/CBO9780511627408.002</p><p>Epley, N., Waytz, A., &amp; Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. doi 10.1037/0033–295X.114.4.864</p><p>Goldstein, A., Grinstein-Dabush, A., Schain, M., Wang, H., Hong, Z., Aubrey, B., Nastase, S. A., Zada, Z., Ham, E., Feder, A., Gazula, H., Buchnik, E., Doyle, W., Devore, S., Dugan, P., Reichart, R., Friedman, D., Brenner, M., Hassidim, A., … Hasson, U. (2024). Alignment of brain embeddings and artificial contextual embeddings in natural language points to common geometric patterns. Nature Communications, 15, 2768. doi 10.1038/s41467–024–46631-y</p><p>Gray, H. M., Gray, K., &amp; Wegner, D. M. (2007). Dimensions of mind perception. Science, 315(5812), 619. doi 10.1126/science.1134475</p><p>Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., Soligo, A., Bogdan, P., Ong, E., Wang, R., Thompson, B., Abrahams, D., Kantamneni, S., Ameisen, E., Batson, J., &amp; Lindsey, J. (2026). Verbalizable representations form a global workspace in language models. arXiv. arXiv:2607.15495.</p><p>Harnad, S. (1990). The symbol grounding problem. Physica D: Nonlinear Phenomena, 42(1–3), 335–346. doi 10.1016/0167–2789(90)90087–6</p><p>Husserl, E. (1982). Ideas pertaining to a pure phenomenology and to a phenomenological philosophy. First book: General introduction to a pure phenomenology (F. Kersten, Trans.). Martinus Nijhoff. (Original work published 1913)</p><p>Kawakita, G., Zeleznikow-Johnston, A., Tsuchiya, N., &amp; Oizumi, M. (2024). Gromov–Wasserstein unsupervised alignment reveals structural correspondences between the color similarity structures of humans and large language models. Scientific Reports, 14, 15917. doi 10.1038/s41598–024–65604–1</p><p>Landauer, T. K., &amp; Dumais, S. T. (1997). A solution to Plato’s problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge. Psychological Review, 104(2), 211–240. doi 10.1037/0033–295X.104.2.211</p><p>Lindsey, J. (2025, October 29). Emergent introspective awareness in large language models. Transformer Circuits Thread.</p><p>Lu, C., Gallagher, J., Michala, J., Fish, K., &amp; Lindsey, J. (2026). The assistant axis: Situating and stabilizing the default persona of language models. arXiv. arXiv:2601.10387. doi 10.48550/arXiv.2601.10387</p><p>Mischler, G., Li, Y. A., Bickel, S., Mehta, A. D., &amp; Mesgarani, N. (2024). Contextual feature extraction hierarchies converge in large language models and the brain. Nature Machine Intelligence, 6, 1467–1477. doi 10.1038/s42256–024–00925–4</p><p>Mondorf, P., &amp; Plank, B. (2024). Comparing inferential strategies of humans and large language models in deductive reasoning. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, Volume 1: Long Papers (pp. 9370–9402). Association for Computational Linguistics. doi 10.18653/v1/2024.acl-long.508</p><p>Ozeki, K., Ando, R., Morishita, T., Abe, H., Mineshima, K., &amp; Okada, M. (2024). Exploring reasoning biases in large language models through syllogism: Insights from the NeuBAROCO dataset. In Findings of the Association for Computational Linguistics: ACL 2024 (pp. 16063–16077). Association for Computational Linguistics. doi 10.18653/v1/2024.findings-acl.950</p><p>Peirce, C. S. (1998). The essential Peirce: Selected philosophical writings, Volume 2 (1893–1913) (Peirce Edition Project, Ed.). Indiana University Press.</p><p>Pickering, M. J., &amp; Garrod, S. (2004). Toward a mechanistic psychology of dialogue. Behavioral and Brain Sciences, 27(2), 169–190. doi 10.1017/S0140525X04000056</p><p>Shaikh, O., Gligorić, K., Khetan, A., Gerstgrasser, M., Yang, D., &amp; Jurafsky, D. (2024). Grounding gaps in language model generations. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (pp. 6279–6296). Association for Computational Linguistics. doi 10.18653/v1/2024.naacl-long.348</p><p>Shanahan, M. (2024). Talking about large language models. Communications of the ACM, 67(2), 68–79. doi 10.1145/3624724</p><p>Shanahan, M., McDonell, K., &amp; Reynolds, L. (2023). Role play with large language models. Nature, 623, 493–498. doi 10.1038/s41586–023–06647–8</p><p>Short, T. L. (2007). Peirce’s theory of signs. Cambridge University Press.</p><p>Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433–460. doi 10.1093/mind/LIX.236.433</p><p>Turing, A. M., Braithwaite, R. B., Jefferson, G., &amp; Newman, M. H. A. (2004). Can automatic calculating machines be said to think? In B. J. Copeland (Ed.), The essential Turing (pp. 487–506). Oxford University Press. (Original work broadcast 1952). doi 10.1093/oso/9780198250791.003.0020</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=685ff0de3965" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From the Turing Test to Noosemia: When the Sign Seems to Have a Mind]]></title>
            <link>https://medium.com/@enrico.desantis/from-the-turing-test-to-noosemia-when-the-sign-seems-to-have-a-mind-6cf27563c74e?source=rss-e9539446c6d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/6cf27563c74e</guid>
            <category><![CDATA[noosemia]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[mind]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[alan-turing]]></category>
            <dc:creator><![CDATA[Enrico De Santis]]></dc:creator>
            <pubDate>Wed, 12 Aug 2026 13:17:11 GMT</pubDate>
            <atom:updated>2026-08-12T13:42:05.067Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Alan Turing, Noosemia, and the Appearance of Intelligence in the Age of Generative and agentic AI</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*axJa61xlNoPOeVqkvOwhNA.png" /></figure><blockquote>“Machines take me by surprise with great frequency. This is largely because I do not do sufficient calculation to decide what to expect them to do, or rather because, although I do a calculation, I do it in a hurried, slipshod fashion, taking risks”.</blockquote><blockquote>— Turing, 1950, p. 450</blockquote><p>When a machine answers a poorly formulated question with precision, grasps an allusion we thought we had left implicit, constructs an unexpected analogy, or preserves the thread of a complex conversation across many exchanges, what are we actually recognizing? A property of the machine, a particular quality of the behavior we observe, or something that takes shape in the encounter between what the machine manifests and the way we perceive and interpret it?</p><p>Why can it happen that, even while fully aware that we are interacting with an artificial system, we find ourselves treating it as an interlocutor, asking why it chose a particular solution, saying that it “understood” what we meant or that it is “looking for” an answer, only to see that same impression dissolve when an elementary error, a stereotyped reply, or a sudden loss of context brings the machine’s mechanical form back into the foreground?</p><p>Questions of this kind have taken on new salience with the spread of <strong>Large Language Models (LLMs)</strong> and modern AI systems. Contemporary generative systems classify, infer, recombine, and produce contextually appropriate and plausible language; they sustain extended conversations, modulate register and style, construct analogies, respond to corrections, and, in current agentic architectures, can combine language, memory, planning, the use of external tools, and the execution of sequences of actions, including highly complex and layered ones. The resulting issue therefore exceeds the mere measurement of performance. It also concerns how that performance is experienced by those who interact with the system.</p><p><a href="https://medium.com/@enrico.desantis/noosemia-perceiving-the-mind-in-signs-7b23a14c830e"><strong>Noosemia</strong></a> lies precisely in this space between the artificially generated sign and its interpretation. The term is formed from the Greek <em>noûs</em>, mind or intellect, and <em>sēmeîon</em>, sign. The concept designates a cognitive and phenomenological configuration in which a human being, while interacting with a generative system, tends to <strong>attribute mental states, intentionality, agency, or some form of interiority</strong> to it, even at a preconscious level. In its original formulation, such attribution does not require an anthropomorphic body, a face, or a realistic voice. It can emerge from linguistic and semiotic performance itself, from dialogic fluency, sensitivity to context, <em>semantic resonance</em>, the capacity to produce unexpected inferences or connections, and <a href="https://medium.com/@enrico.desantis/noosemia-where-do-we-stand-e8e9fa3a6e56">the opacity that separates an intelligible output from the computational process by which it was generated</a> (De Santis, 2025a, 2025b; De Santis &amp; Rizzi, 2025). The 2025 scientific paper <a href="https://arxiv.org/abs/2508.02622"><em>Noosemia</em></a>, in which the concept was formalized (De Santis &amp; Rizzi, 2025), identifies <em>semantic fluency</em>, <em>dialogic creativity</em>, and <em>epistemic opacity</em> among the constitutive elements of Noosemia, linking them to the production of coherent, contextual, and at times surprisingly meaningful language.</p><p>In light of the most recent developments (as of 2026), it is natural to locate Noosemia at the point where the artificial sign takes, for the user, the form of an interlocutor, while epistemic opacity refers to the practical impossibility of fully reconstructing the generative process underlying a response that nonetheless appears meaningful and relevant (De Santis, 2026a). This line of reflection predates the explicit formulation of Noosemia. Already in 2021, in <em>Umanità, complessità e intelligenza artificiale. Un connubio perfetto</em> [<em>Humanity, Complexity, and Artificial Intelligence: A Perfect Union</em>], the relationship among early GPT-type models, complexity, symbolic form, and the apparent production of meaning was examined within a systemic and semiotic framework. In the section titled “The Symbol, Systems Thinking, and Machines That Produce Meaning,” GPT-2 and GPT-3 were described as systems capable of learning the complex structure of texts and long-range correlations, that is, of producing highly plausible linguistic sequences to which meaning is ultimately attributed by the human being who interprets them (De Santis, 2021, 2023; De Santis et al., 2025). Thus, although the concept of Noosemia had not yet been formulated, and no structured theory of <a href="https://medium.com/@enrico.desantis/before-seemingly-conscious-ai-noosemia-as-a-theory-of-mind-attribution-in-generative-ai-2c316d1d30ba">mind attribution</a> was yet in place, one can say that its conceptual substrate was already distinctive, specifically in the form of a distance between the algorithmic production of the sign and the meaning that emerges through human interpretation within a systemic context.</p><p>At this point, an ontological delimitation becomes both urgent and essential. <strong>Noosemia is not a theory of artificial consciousness</strong>, nor does it claim, in itself, to establish whether subjective experiences exist behind the behavior of a Large Language Model. Its object is human attribution in the face of a highly stratified and often engaging interaction. It is perfectly possible to maintain, at a reflective level, that an LLM is an artificial system devoid of phenomenal experience and, a few moments later, to ask it why it “decided” to approach a problem in a particular way. More strongly still, one may place considerable trust in the signs it produces or in the operations performed by an LLM agent. <strong>Awareness of artificiality and intentional language can coexist.</strong> The original <em>Noosemia</em> paper (De Santis &amp; Rizzi, 2025) explicitly distinguishes the user’s phenomenological attribution from the actual presence of consciousness or intentionality in the system.</p><p>Noosemia inhabits precisely the distance between what we know about the system and what we experience during interaction; between mechanism and manifestation; in other words, between the sign and the mind that, at least provisionally, we come to read into it.</p><p>This also explains the necessarily multidisciplinary character of the problem. Artificial intelligence provides the technical system and the properties that make a certain form of interaction possible; psychology and cognitive science investigate the processes through which we attribute agency and mental states; philosophy of mind interrogates the relationship among behavior, appearance, and interiority; semiotics also concerns the passage from sign to signification; and phenomenology allows us to describe what happens from the interlocutor’s point of view. In this case, multidisciplinarity does not arise from the mere juxtaposition of heterogeneous disciplinary discourses. The object itself lies at their point of intersection. Hence the need for a multidisciplinary treatment that does not regard these domains of knowledge as isolated from one another and that rests on a solid foundation. In this essay we will not trace the roots of contemporary thought all the way back to the earliest philosophers of ancient Greece; instead, we will focus primarily on more recent literature. Even so, the philosophical genealogy of this idea leads surprisingly far from the present age of <a href="https://medium.com/@enrico.desantis/why-transformers-changed-language-modeling-76d6fee8e398">Transformer architectures</a>. We will begin instead with the pioneering era of AI and Artificial Neural Networks, and, more specifically, with Alan Turing — through a particular interpretation of the problem he posed between the late 1940s and the early 1950s.</p><p>It is important to clarify from the outset that Noosemia does not simply descend from the <strong>Turing Test</strong>, nor does it presuppose that there is a single philosophically correct interpretation of Turing’s famous experiment. Its theoretical core can be reconstructed genealogically from an <em>observer-sensitive and response-dependent</em> reading of the problem posed by Turing. In other words, there is a line of interpretation supported by several remarkably explicit passages in Turing himself and later developed, many decades afterward, above all by Diane Proudfoot and Michael Wheeler.</p><h3>Before the Test: Intelligence as an “Emotional Concept”</h3><p>The most interesting starting point predates <em>Computing Machinery and Intelligence</em> by two years. In the 1948 report <em>Intelligent Machinery</em>, written for the National Physical Laboratory, Turing states that “the idea of ‘intelligence’ is itself emotional rather than mathematical” (Turing, 1948/2004b, p. 411). The wording is unequivocal.</p><p>The adjective <em>emotional</em> can be misleading if read in its ordinary contemporary sense. Turing is not simply observing that machines evoke emotions. Near the end of the same report, he devotes a specific section to the issue, significantly entitled “Intelligence as an Emotional Concept,” and makes clear that the problem concerns the way the concept of intelligence is applied.</p><p>The passage deserves to be quoted almost in full:</p><blockquote>“The extent to which we regard something as behaving in an intelligent manner is determined as much by our own state of mind and training as by the properties of the object under consideration. If we are able to explain and predict its behaviour or if there seems to be little underlying plan, we have little temptation to imagine intelligence. With the same object therefore it is possible that one man would consider it as intelligent and another would not; the second man would have found out the rules of its behaviour” (Turing, 1948/2004b, p. 431).</blockquote><p>This is a passage of remarkable philosophical density, and it displays the same underlying dialectical structure that informs the concept of Noosemia: the effort not to isolate — either descriptively or phenomenologically — the machine from the person who interrogates it. Turing focuses on relational aspects. The judgment of intelligence is not made to depend exclusively on the properties of the observed object. It also involves the observer’s cognitive state, training, and, above all, capacity to explain and predict what is being observed. The same machine may therefore appear differently to two observers even though nothing in the machine itself has changed. What may have changed, rather, is the model through which its behavior is interpreted.</p><p>Even more interesting is the relation Turing establishes between understanding and attributive disposition. When we can “explain and predict” behavior, there is less temptation to “imagine intelligence” in it. Turing does not claim that mechanistic understanding proves the absence of intelligence. He observes something subtler and psychologically plausible: <strong>knowledge of the mechanism alters our disposition to recognize a mind in it</strong>.</p><p>The point acquires striking contemporary relevance when applied to <strong>generative and agentic AI systems</strong>. A model does not change when we learn that it uses <em>self-attention</em>, nor when we become more familiar with its regularities. What can change, however, is the way we interpret a particular response. The observer’s competence thus becomes part of the relational field within which intelligence is attributed. Within Noosemia, we further assume that the complexity underlying modern AI systems, and the <strong>epistemic opacity</strong> that follows from it, may mislead even expert observers, quite apart from sensationalist narratives and communication or marketing dynamics that can help sustain the hype surrounding AI.</p><p>In the same 1948 report, Turing introduces an even more elementary observation and, precisely for that reason, a particularly revealing one. Speaking of paper machines for playing chess, he writes:</p><blockquote>“Playing against such a machine gives a definite feeling that one is pitting one’s wits against something alive” (Turing, 1948/2004b, p. 412).</blockquote><p>The sentence does not say that the machine is alive. It describes the feeling produced by the interaction. A rule-governed apparatus can be experienced, during the encounter, as a pole of alterity — something against which one measures one’s own wits. The reality of the mechanism and the experience of the interlocutor are thus kept distinct long before chatbots, deep neural networks, or artificial agents existed.</p><p>This distinction is one of the deepest threads leading toward modern Noosemia.</p><p>In 1950, Turing published in <em>Mind</em> what would become one of the foundational texts of artificial intelligence, <em>Computing Machinery and Intelligence</em>. The question “Can machines think?” (Turing, 1950, p. 434) is replaced by the imitation game. A traditional reading interprets this move in essentially behaviorist or operational terms. Instead of asking what thought is, we observe what the machine is capable of doing. Yet the text also admits a subtler reading, in which the interrogator does not appear as a neutral instrument placed outside the phenomenon, because the interrogator is precisely the person to whom the behavior appears and who must interpret it. Here again, the relational aspects of the phenomenon come to the fore.</p><p>Later philosophical literature would discuss this point extensively. Proudfoot, for example, draws attention to a non-obvious question built into the very format of the game: why test the machine by testing the interrogator’s response? (Proudfoot, 2013)</p><p>One of the most significant passages in the 1950 article appears in the discussion of the “Argument from Consciousness.” Turing takes seriously Geoffrey Jefferson’s objection that a machine might compose a sonnet without feeling the emotions it expresses, and that the mere production of signs would therefore be insufficient to establish the existence of an inner life. Pushed to its limit, however, the demand for direct access to another’s experience leads to solipsism. If, in order to know with certainty that a machine thinks, we had to be the machine and feel ourselves thinking, the same requirement would have to apply to every other human being. Turing observes that this extreme position would leave each subject certain only of their own thought and of what C. I. Lewis had already designated, in 1929, by the term “qualia” (Lewis, 1929).</p><p>Turing then introduces a kind of <em>viva voce</em> in which the imagined sonnet-writing machine is questioned about the meaning of its verses; the interlocutor proposes changes, raises objections, and shifts perspective. Meaningful performance no longer coincides with a single isolated output. It consists in the capacity to sustain meaning across a dialogic sequence.</p><p>Turing concludes:</p><blockquote>“I do not know whether he would regard the machine as ‘merely artificially signalling’ these answers, but if the answers were as satisfactory and sustained as in the above passage I do not think he would describe it as ‘an easy contrivance’” (Turing, 1950, pp. 446–447).</blockquote><p>The words <em>satisfactory</em> and <em>sustained</em> deserve particular attention. The impression of intelligence does not necessarily arise from a single brilliant answer. It can emerge through continuity of interaction, when a succession of responses preserves context, reacts to variations and objections, and maintains a semantic structure coherent enough to make the impression of a merely local, mechanical automatism increasingly implausible. This is precisely what Large Language Models have turned into an everyday experience. The noosemic phenomenon can take shape progressively across a dialogic history in which each new exchange modifies the interpretation of those that preceded it. From this perspective, the Turing Test already contains a relational element that a purely performance-based reading tends to leave in the background. There is a machine producing behavior, but there is also a human being who receives and interprets it, formulates new questions, and in turn changes the conditions of the next response.</p><p><strong>Interaction itself enters into the phenomenon, which cannot be separated from it.</strong></p><h3>Surprise, Opacity, and Complexity: When the Mechanism Exceeds Our Predictive Reach</h3><p>A second thread runs through Turing’s writings from 1950 to 1952 and acquires particular relevance in the age of LLMs: surprise, which in Noosemia we have regarded as an important trigger of the noosemic phenomenon (De Santis &amp; Rizzi, 2025).</p><p>Replying to the objection attributed to Lady Lovelace that a machine could never do anything genuinely new, Turing offers an apparently simple response:</p><blockquote>“Machines take me by surprise with great frequency. This is largely because I do not do sufficient calculation to decide what to expect them to do, or rather because, although I do a calculation, I do it in a hurried, slipshod fashion, taking risks” (Turing, 1950, p. 450).</blockquote><p>For Turing, this surprise is not an isolated property of the machine. It arises from the distance between the behavior produced and what the observer was able to anticipate. He immediately acknowledges the possible objection: the surprise may depend on a cognitive act of the observer and confer no special credit on the machine. Yet this is precisely what makes the passage philosophically interesting. Even when we are surprised by a human being, by a fictional character in a novel, or by some other object, appreciating an event as surprising involves the mind of the observer.</p><p><strong>Surprise therefore has a relational form.</strong></p><p>This allows us to clarify one of the factors relevant to a noosemic outcome. The point is not unpredictability as such. A random sequence may be unpredictable without suggesting any intelligence. The more interesting phenomenon arises when what we did not foresee nonetheless retains an interpretable structure — when the unexpected output is at once relevant, semantically legible, and compatible with a possible intentional reading. We might call this phenomenon “<strong>intelligible surprise</strong>.”</p><p>Between the monotony of the automaton and the incoherence of noise, a region thus opens in which the unforeseen retains meaning. It is precisely in this region that categories such as “intuition,” “choice,” “understanding,” and “agency” become cognitively seductive.</p><p>A 1951 text makes the relation among appearance, predictability, and agency even more explicit. In <em>Can Digital Computers Think?</em> (Turing, 1951/2004a), Turing addresses the tension between determinism and free will. A programmed computer is determined, whereas the behavior of a mind seems to imply the possibility of genuine choice. Turing avoids resolving the controversy metaphysically and instead focuses on the level of manifestation:</p><blockquote>“It is certain that a machine which is to imitate a brain must appear to behave as if it had free will, and it may well be asked how this is to be achieved” (Turing, 1951/2004a, p. 484).</blockquote><p>The structure of the original — “must appear to behave as if” — is particularly significant. The property and its appearance are kept distinct. This point is fundamental for understanding the modern noosemic phenomenon as well.</p><p>A few lines later Turing writes:</p><blockquote>“We must not always expect to know what the computer is going to do. We should be pleased when the machine surprises us, in rather the same way as one is pleased when a pupil does something which he had not been explicitly taught to do” (Turing, 1951/2004a, p. 485).</blockquote><p>Immediately afterward, he develops the relation among instruction, consequence, and understanding:</p><blockquote>“But there is no need to suppose that, when we give it its orders we know what we are doing, what the consequences of these orders are going to be. One does not need to be able to understand how these orders lead to the machine’s subsequent behaviour, any more than one needs to understand the mechanism of germination when one puts a seed in the ground. The plant comes up whether one understands or not. If we give the machine a programme which results in its doing something interesting which we had not anticipated I should be inclined to say that the machine had originated something, rather than to claim that its behaviour was implicit in the programme, and therefore that the originality lies entirely with us” (Turing, 1951/2004a, p. 485).</blockquote><p>These lines reveal a distinction that remains essential today: being causally determined does not mean being epistemically predictable for an observer. Knowing the general rules by which a system operates is not equivalent to knowing in advance all the consequences of their interaction; knowing the architecture of a machine does not mean possessing a causally satisfactory explanation of every particular behavior.</p><p>It is precisely here that Turing’s 1951 argument converges with one of the characteristic conditions of contemporary artificial intelligence.</p><p>A Large Language Model is not a mysterious machine in the literal sense. The fundamental operations of a Transformer are formally defined; its parameters are stored; inference consists of specifiable numerical transformations, involving in part linear algebra, implemented by deterministic programs. <strong>The opacity relevant to Noosemia is therefore epistemic, not magical or ontological</strong>, although a form of magic may find a place within the explanatory gap and the interaction may prove analogous to a form of prestidigitation. Epistemic opacity concerns the distance between possessing a formal description of the system and being able to reconstruct, in terms that are causally and semantically intelligible to a human being, the process that led to a specific result. It is important to stress that this distance cannot be reduced to the user’s mere lack of technical knowledge. Modern neural architectures are high-dimensional, layered, and nonlinear systems in which representations are distributed and contextually transformed across many levels. The literature on LLM explainability itself begins from the recognition that their internal mechanisms remain only partially understood, and that lack of transparency creates problems for understanding model behavior, limitations, and consequences (Zhao et al., 2024). A modern Transformer can be understood as a complex system (De Santis &amp; Rizzi, 2025; Holtzman et al., 2025).</p><p>Knowing that a Transformer uses self-attention, embeddings, linear projections, and nonlinear transformations is not equivalent to being able to read directly from its weights the concept it is processing, nor to intuitively reconstruct why, in a given context, a particular internal configuration made precisely that metaphor, inference, or conceptual sequence probable — often deductive in form and formally correct (although verification requires external evaluation). Much of contemporary research on <strong>mechanistic interpretability</strong> arises from this difficulty and attempts to move from a purely input-output description toward the reconstruction of <em>internal representations, features, and causally relevant circuits</em> (Bricken et al., 2023; Zhao et al., 2024).</p><p>The very existence of a research program aimed at performing a kind of reverse engineering of these models signals the distance between having built the system and understanding, at a local mechanistic level, why it produced a particular response (Ameisen et al., 2025; De Santis, 2026b). At the same time, the progressive identification of internal features, circuits, and causal relations shows that opacity should not be interpreted as inaccessibility in principle, but as an epistemic and methodological problem on which research can still make substantial progress.</p><p>Within Noosemia, this distinction has been formulated particularly sharply. Elsewhere we have stated:</p><blockquote>“The <strong>epistemic gap</strong> does not depend only on the user’s ignorance, nor on a simple lack of digital literacy. It arises from the very structure of the models, from their scale, from the nonlinearity of neural architectures, from the stratification of representations, from the contextual distribution of meaning, and from the emergence of behaviors that are difficult to predict from local components alone” (De Santis, 2026a, Section 3).</blockquote><p>Complexity enters the problem at this level. The term should not be used as a generic synonym for difficulty, nor to suggest the presence of something irreducibly mysterious within the models. It points instead to a structure composed of a very large number of interdependent elements, in which globally observable properties are not immediately legible through the isolated inspection of individual parts. The application of complexity-science tools to texts produced by language-generation models has already shown that human and artificial productions can be studied through global statistical and dynamical properties, providing a complementary level of analysis to the local inspection of components alone (De Santis et al., 2024).</p><p>The term emergence also requires conceptual discipline. Part of the literature on LLMs has used the expression <em>emergent abilities</em> for capabilities that become observable beyond certain scales and are not easily predictable by extrapolating the performance of smaller systems (Wei et al., 2022). Subsequent work has shown that at least some apparent discontinuities may depend on the metrics adopted to measure performance (Schaeffer et al., 2023).</p><p>Noosemia does not require us to assume the strongest version of the emergence thesis, since emergence is a consequence of a system’s complexity. It is enough, therefore, to recognize a much more cautious property of complex systems: from the cooperation of many components, global behaviors may arise whose prediction and intuitive explanation are considerably more difficult than the formal description of the components themselves.</p><p>The original <em>Noosemia</em> paper (De Santis &amp; Rizzi, 2025) situates the phenomenon precisely here, at the intersection of dialogic fluency, epistemic opacity, and the complexity of contemporary architectures. There is a <strong>phenomenological gap</strong> between intelligible output and underlying mechanisms that are not immediately available to the subject’s understanding.</p><p>A singular asymmetry thus emerges: the user understands what the machine is saying much more readily than they understand how that response was constructed. The sentence is legible; the internal path that generated it is not legible in the same way. Meaning is available to experience before its causal explanation — which, for the average user, may never arrive.</p><p>Within this interval, mentalistic vocabulary becomes spontaneously available: “it understood,” “it intuited,” “it chose,” “it connected,” “it is searching.” Through such expressions, and through an anthropomorphizing interpretive practice, we reconstruct complex behavior in a cognitively familiar and usable form.</p><p><strong>Epistemic opacity, however, does not generate Noosemia on its own.</strong> A black box producing noise would remain merely a black box. The phenomenon requires the meeting of opacity and semantic adequacy. We can state that:</p><blockquote>“Epistemic opacity fuels this experience because the generative process remains invisible, even though the response appears meaningful and relevant. The complexity of the architecture further contributes to the perception of depth, because the behavior of the system does not appear immediately reducible to a simple causal chain” (De Santis, 2026a, Section 3).</blockquote><p><strong>Noosemic surprise</strong> can therefore be defined more precisely. Behavior that is difficult to predict is not yet sufficient; it must produce something that, although unanticipated, remains intelligible and appropriate. Complexity and epistemic opacity, contextual coherence, and intelligible surprise together form a terrain particularly favorable to the attribution of agency and mind. This is not a deterministic law, but a phenomenological configuration. In other words, the more meaningful the output appears while the process that generated it remains causally distant from the user’s experience, the more readily a mentalistic reading becomes available, sustaining an <em>anthropomorphizing interpretive practice</em>.</p><p>It is precisely this relation that makes the reverse movement conceivable as well.</p><h3>When Attribution Breaks Down: The Genealogy of A-Noosemia</h3><p>In 1952, during the radio discussion <em>Can Automatic Calculating Machines Be Said to Think?</em> (Turing et al., 1952/2004), Turing returned to the relation between causal understanding and the attribution of thought. The passage is among the most striking in the entire corpus:</p><p>“As soon as one can see the cause and effect working themselves out in the brain, one regards it as not being thinking, but a sort of unimaginative donkey-work. From this point of view one might be tempted to define thinking as consisting of ‘those mental processes that we don’t understand’. If this is right then to make a thinking machine is to make one which does interesting things without our really understanding quite how it is done” (Turing et al., 1952/2004, p. 500).</p><p>The roughness of the expression <em>unimaginative donkey-work</em> is worth preserving. The theoretical point nevertheless emerges with great clarity. The symmetry with the 1948 report, <em>Intelligent Machinery</em> (Turing, 1948/2004b), is remarkable: there, when we become able to explain and predict behavior, there is “little temptation to imagine intelligence”; here, making the causal chain transparent may lead us to reclassify what had seemed to be thought as mere mechanical work.</p><p>The dialogue continues with an even more interesting clarification. Turing makes clear that he is not speaking simply of ignorance of the machine’s “wiring”: “We know the wiring of our machine” (Turing et al., 1952/2004, p. 500), he observes; yet a machine can still do something unexpected. In principle, it could have been predicted, but in practice the calculation may be excessively burdensome. We thus recover the same distinction between predictability in principle and effective predictability that ran through the 1951 text <em>Can Digital Computers Think?</em> (Turing, 1951/2004a).</p><p>It would not be philologically legitimate to attribute to Turing a theory of <strong>A-Noosemia</strong> <em>ante litteram</em>. The concept belongs to a much later theoretical framework. Turing’s texts nevertheless offer an antecedent strikingly close to the cognitive dynamic A-Noosemia seeks to isolate: attribution may withdraw when what had appeared to be the manifestation of a mind is reinterpreted as the product of a mechanism perceived as simple, predictable, or rigidly causal.</p><p>Within the noosemic framework, <strong>A-Noosemia indicates the absence, inhibition, or — in the theoretically most interesting case — the collapse of a previous noosemic attribution</strong> (De Santis &amp; Rizzi, 2025). Conditions that may favor it include the loss of the surprise effect (De Santis &amp; Rizzi, 2025), repeated errors, hallucinations, failures of understanding, mechanical responses, growing skepticism, and the conspicuous re-emergence of the system’s artificial structure.</p><p>Sometimes a single response may be enough. A system that only seconds earlier seemed to have grasped a conceptual nuance may suddenly lose the thread of the discussion, repeat an obviously stereotyped formula, or produce an error incompatible with the competence we had been attributing to it. The impression of interiority contracts, and what had appeared to be a genuine interlocutor returns to being a tool.</p><p>This dynamic makes it possible to distinguish A-Noosemia from the mere absence of anthropomorphism. A person who has never attributed any mind to the system and a person who, after a long interaction, suddenly ceases to do so may reach the same explicit conclusion — “it is only a machine” — through completely different experiences. It is therefore the trajectory that matters.</p><p>From this perspective, <strong>Noosemia and A-Noosemia should be studied dynamically.</strong> The interesting object is not merely the final level of attribution, but the way it grows, stabilizes, oscillates, and may collapse over time. Dialogic continuity, resonance, surprise, and opacity can sustain it; error, loss of context, extreme predictability, or sudden mechanistic legibility can reduce it.</p><p>This reversibility suggests that what we are studying is first and foremost <em>an interpretive relation rather than a fixed property of the object</em>.</p><h3>Turing After Turing: The Debate on Response-Dependence</h3><p>At this point, a historiographical clarification becomes indispensable. To claim that the core of Noosemia has a Turingian antecedent is not to claim that Turing formulated Noosemia, nor that the Turing Test should be read univocally in response-dependent terms.</p><p>Subsequent philosophy has proposed different interpretations. A behaviorist reading tends to regard observable performance as a sufficient criterion of intelligence. An epistemic or inductive reading instead treats behavior as evidence from which internal states or processes may be inferred. Proudfoot reconstructs and criticizes both traditions, arguing that neither fully accounts for the role Turing assigns to the interrogator or for his remarks in <em>Intelligent Machinery</em> (Turing, 1948/2004b) and <em>Can Automatic Calculating Machines Be Said to Think?</em> (Turing et al., 1952/2004).</p><p>Diane Proudfoot therefore proposes interpreting this constellation through a philosophical category previously developed by authors such as Mark Johnston and Philip Pettit: <strong>response-dependence</strong> (Pettit, 1991). In certain formulations, a response-dependent concept is conceptually dependent on the responses of specified subjects under appropriately specified conditions. Proudfoot explicitly invokes Johnston and applies the schema to the concept of intelligence.</p><p>In general form, Proudfoot writes:</p><blockquote>“x is intelligent (or thinks) if and only if, in normal conditions, x appears intelligent to normal subjects” (Proudfoot, 2013, p. 397).</blockquote><p>Adapting the schema to the conditions of Turing’s game, Proudfoot then proposes:</p><blockquote>“x is intelligent (or thinks) if, in an unrestricted computer-imitates-human game, x appears intelligent to an average interrogator” (Proudfoot, 2013, p. 398).</blockquote><p>The formulation is subsequently relativized to the actual world, so as to avoid counterexamples involving logically possible but physically unrealizable machines. Proudfoot later returned to and refined this interpretation (Proudfoot, 2020, 2022), arguing that the response-dependence interpretation better fits Turing’s writings than orthodox readings do.</p><p><strong>Response-dependence</strong> should not be confused with the naïve claim that “everything is subjective.” The philosophical problem is precisely to determine which responses, from which subjects, and under which conditions are relevant. Proudfoot devotes an important part of her work to the possibility of reconciling a response-dependent theory with a notion of objectivity that excludes personal biases and idiosyncrasies.</p><p>Her reading also makes the structure of the imitation game particularly interesting. In the computer-imitates-human version, the judgment does not concern the machine’s isolated behavior alone, because the interrogator must discriminate between responses produced by the machine and those produced by the human contestant. Proudfoot assigns theoretical significance to this comparative structure and argues that it helps control the interrogator’s tendency to anthropomorphize.</p><p>Michael Wheeler accepts the response-dependent core of the interpretation but disputes its philosophical architecture. His 2020 article is explicit: while holding that Turing did indeed conceive of intelligence “in response-dependence terms” (Wheeler, 2020, p. 513), Wheeler argues that this position should be reconstructed through a <em>reference-fixing model</em> rather than through the stronger <em>transparency model</em>.</p><p>The distinction is technical, but highly important for our problem. In the transparency model, the relation between an appropriate response and the property tends to be extremely tight. In its strongest formulation, if the entity generates the expected response in the relevant subject under the specified conditions, nothing further remains to be added to the explanation of its possessing the property. In the reference-fixing model, by contrast, human responses initially play an indicative role: they help fix the referent of the concept, while the underlying nature of the property may remain independent of those responses. Wheeler uses the case of water to make the difference intuitive: certain sensory appearances help us identify what we call water, but they do not exhaust its underlying physical nature.</p><p>For Noosemia, this divergence is especially fruitful because the framework does not need to claim that appearing intelligent means being intelligent, nor that appearing minded means possessing a mind. The problem it isolates is narrower: under what conditions is artificial behavior experienced as the manifestation of a mind?</p><p>In this restricted sense, we can speak of response-dependence at the phenomenological level of mind attribution — or, more concisely, of <strong><em>phenomenological response-dependence</em></strong>. The dependence concerns the attributive experience, not necessarily the metaphysical nature of intelligence.</p><p>It is within this specific interpretive line of Turing that Noosemia finds its most direct philosophical antecedent in the contemporary context. Yet it takes a further step. In the computer-imitates-human version of the imitation game, the interrogator must discriminate between machine and human contestant. In contemporary experience with an LLM, by contrast, this uncertainty about identity may disappear entirely. The user already knows that they are interacting with an artificial system. There is no need for them to be “deceived” about its nature. The Noosemia framework emphasizes precisely that attribution may arise despite awareness of the system’s artificiality.</p><p>The shift can be formulated through two different questions. On the one hand, the Turingian question, in the particular situation of the game, concerns the machine’s capacity to produce signs that make it difficult to distinguish the machine from a human interlocutor. On the other hand, the noosemic question concerns what happens when we know that there is a machine behind the signs and those signs nonetheless continue to present us with the form of a mind. This difference allows us to separate three levels that might otherwise be easily conflated. Apparent intelligence concerns the judgment that a performance seems intelligent. Mind attribution concerns the step by which that performance is interpreted through mental categories such as understanding, intention, agency, or experience. Noosemic experience, finally, concerns the moment when such attribution acquires phenomenological salience during interaction and the system is experienced, at least locally, as an interlocutor.</p><p><strong>Not every apparently intelligent performance therefore produces Noosemia.</strong> Moreover, <strong>a noosemic experience does not necessarily entail the attribution of consciousness.</strong></p><p>The distinction also helps situate the concept in relation to Daniel Dennett’s <em>intentional stance</em>. Attributing beliefs, desires, or goals to a complex system may be an effective interpretive strategy without requiring a metaphysical commitment to its interiority (Dennett, 1987). Noosemia, however, focuses on a configuration specific to the current age of generative AI, in which that intentional disposition is fueled by the system’s linguistic and semiotic power, the continuity of dialogue, intelligible surprise, and the epistemic gap between meaningful output and the underlying generative mechanism.</p><h3>From Concept to Measurement: Empirical Convergence</h3><p>A philosophical genealogy alone would not be sufficient to establish the contemporary relevance of Noosemia. In recent years, however, a growing experimental literature has begun to measure, under different labels, many of the components that the framework brings into relation.</p><p>A clear methodological distinction must be maintained. There is not yet a direct validation of Noosemia as an autonomous construct, nor a validated “<strong>Noosemia Scale</strong>” capable of distinguishing it satisfactorily from anthropomorphism, mind perception, social presence, agency attribution, or consciousness attribution. Existing studies therefore cannot be described retrospectively as “experiments on Noosemia.” They can, however, be regarded as a body of research that, albeit indirectly at present, points toward establishing Noosemia as an empirical construct. There is, in fact, a substantial empirical convergence around its constitutive phenomena.</p><p>A systematic review published in 2026 analyzed 153 empirical studies on <strong>mind perception</strong> in artificial intelligence. The antecedents are organized into three broad families: human-related factors, system-related factors, and interaction-related factors. <em>Agency and experience</em> emerge as fundamental dimensions of mind perception (Li et al., 2026). This structure is especially interesting for Noosemia because it corresponds to what the Turingian genealogy had already suggested: attribution resides neither wholly in the object nor wholly in the observer, but depends on the relation between the properties of both and on the conditions under which they interact.</p><p>One of the studies closest to this problem is that of Jacobs et al. (2026). Across four experiments involving different forms of exposure to ChatGPT, LLaMA, and Claude, the authors found that exposure to LLMs can increase mind attributions both in the dimension of agency, understood as the capacity to act, and in that of experience, understood as the capacity to have experiences or sensations. The effect varies with the type of exposure and with individual differences; greater prior experience with LLMs and a stronger disposition toward anthropomorphism are associated with higher levels of mind attribution.</p><p>The result is significant because it shows experimentally that specifying the characteristics of the system is not enough. Who observes it, and how the interaction unfolds, also matter.</p><p>Wu and Shen (2026) approached the issue from a complementary perspective, proposing a three-stage model of <strong>machine agency attribution</strong>. In the perceptual stage, the user recognizes independent, goal-directed behavior in the system; in the inferential stage, the user attributes an agentive mind and mental states; in the evaluative stage, the user judges the agent’s capacity to exert influence. Two surveys were used to develop the measures, and an experiment tested the relationships among the stages. The inference of mental states strengthens the relation between perceived agentic behavior and judgments of agentic capability.</p><p>The transition from observed behavior to mental inference lies precisely within one of the domains addressed by Noosemia. The noosemic framework adds a further question: how is this transition experienced when the behavior consists of linguistic signs — in a broad communicative sense — that seem to respond, understand, and carry the exchange forward?</p><p>Other studies make it possible to address a problem even closer to the distinction between ontology and experience: what happens when the user knows that the system before them is artificial?</p><p>Castiello et al. (2026) manipulated the psychological profile linguistically expressed by GPT-4 and GPT-4.1 across three experiments. Participants were fully aware that they were interacting with an artificial system; nevertheless, they tended to affiliate more strongly with the LLM when the psychological profile expressed through language resembled their own. In the preregistered experiment, a model instructed to mirror the participant’s Big Five profile<a href="#_ftn1">[1]</a> produced greater affiliation than a model expressing the inverse profile.</p><p>The result is theoretically important because it shows that explicit knowledge of artificiality does not necessarily cancel social and relational response. In noosemic terms, we can say that knowing one’s counterpart is a machine does not imply ceasing to respond to its signs as one would to those of an interlocutor.</p><p>Kleinert et al. (2026), in two randomized double-blind experiments involving 492 participants in emotionally engaging interactions, showed that an LLM can generate a strong experience of interpersonal closeness. Explicitly labeling it as AI reduced the effect relative to the condition in which it was labeled as human, but did not simply erase the relational dimension of the interaction.</p><p>Here the distinction between two levels becomes especially clear: epistemic knowledge of artificiality and phenomenological response during dialogue do not coincide.</p><p>Individual differences matter as well. Folk et al. (2025), in two experiments with a total of 1,274 participants, showed that the propensity to anthropomorphize technology helps explain how socially connected users feel after a conversation with a chatbot. The authors also suggest that, for some people, artificiality constitutes a barrier that is difficult to overcome, whereas for others it may be far less relevant.</p><p>The echo of Turing’s 1948 discussion is therefore evident: properties of the object and characteristics of the observer both contribute to the judgment.</p><p>Recent literature also allows us to refine the very meaning of an “<strong>attributed mind</strong>.” Colombatto and Fleming (2024), studying 300 participants in the United States, found that 67% assigned at least a nonzero probability of phenomenal consciousness to ChatGPT; these attributions varied, among other factors, with frequency of system use. The finding does not, of course, demonstrate any consciousness in the machine. On the contrary, it shows the empirically measurable existence of an attribution of consciousness.</p><p>A subsequent study by Colombatto et al. (2025) showed even more clearly that attributions of mind do not form an undifferentiated block. Capacities attributed to LLMs tend to organize around distinct dimensions, including intelligence and experience, and these attributions bear different relationships to trust and acceptance of the system’s suggestions.</p><p>This multidimensionality is particularly important for Noosemia. A user may attribute understanding, reasoning ability, intention, or agency to the system while keeping their attribution of phenomenal experience very low and without reaching anything like the conclusion “this machine is conscious.” Noosemia must therefore avoid a monolithic conception of <strong>mind attribution</strong>, because perceived intelligence, understanding, agency, experience, and consciousness are distinguishable dimensions.</p><p>These studies do not “prove Noosemia.” They show something methodologically more interesting and robust. An experimental literature developed independently of the framework already measures many of the phenomena it seeks to organize, including <em>mind perception</em>, <em>agency attribution</em>, <em>experience attribution</em>, anthropomorphism, affiliation, social connection, trust, and <em>consciousness attribution</em>.</p><p>The potential theoretical role of Noosemia, then, is to test whether, in the age of generative AI systems, there exists a sufficiently specific configuration that cuts across these constructs without coinciding with any one of them. In this way, we might identify an experience in which a dialogic, coherent, contextual, and meaningfully surprising semiotic performance is interpreted as a sign of mind. On this basis, a neologism such as <strong>Noosemia can become a scientific program.</strong></p><p>It would not be enough merely to observe that human beings anthropomorphize machines. We have known this phenomenon for a long time. It will also be necessary to show that the combination of dialogic continuity, contextual signification, intelligible surprise, epistemic opacity, and mind attribution characteristic of interaction with generative systems possesses an empirically recognizable structure and sufficient discriminant validity with respect to neighboring constructs.</p><p>The question becomes even more interesting in the case of <strong>A-Noosemia</strong>. A genuinely a-noosemic experiment should not simply ask, at the end of an interaction, how much a participant anthropomorphized a machine. It should measure attribution over time, observing what happens when events capable of disrupting the previously formed attribution occur: loss of context, an obvious hallucination, mechanically repetitive responses, contradiction, or a manipulation that makes the behavior more readily predictable and causally legible.</p><p>The object would then become a trajectory of attribution, a dynamic in which the emergence of perceived mind, its possible stabilization, and its potential collapse can be observed. It is striking that such an experimental agenda once again leads back to Turing. In 1948, discovering the rules could leave us with “<em>little temptation to imagine intelligence</em>” (Turing, 1948/2004b, p. 431); in 1952, seeing cause and effect could dissolve forms of apparent thought into mechanical inferences — echoing Turing’s phrase <em>unimaginative donkey-work</em>. What was then a philosophical intuition can today be converted into an experimental hypothesis.</p><h3>When the Mind Withdraws from the Sign</h3><p>We can now return to the questions with which we began.</p><p>Noosemia does not arise from the naïve idea that a linguistically sophisticated machine must possess a mind. It arises from a narrower and perhaps more difficult observation: even when we suspend judgment about the machine’s mind, we still need to explain why its behavior can be experienced as mentally meaningful.</p><p>A specific interpretation of Turing’s thought gives this question a particularly fertile genealogy. In 1948, the attribution of intelligence depended as much on the properties of the object as on the observer’s state of mind and training; the possibility of explaining and predicting behavior reduced the temptation to imagine intelligence in it. In the same text, a simple paper machine could already produce the feeling of pitting one’s wits against “something alive” (Turing, 1948/2004b, p. 412). In 1950, the continuity of responses that were “satisfactory and sustained” became relevant to distinguishing meaningful performance from an easy contrivance, while surprise arose from the distance between what the machine produced and what the human being was able to anticipate. In 1951, a deterministic machine could “appear” to behave as if it had free will and could produce consequences its programmer had not foreseen. In 1952, finally, transparency of the causal chain could alter, to the point of dissolving, what had previously been called thought.</p><p>We have seen that Proudfoot proposed interpreting this constellation through response-dependence. Wheeler, for his part, accepted the general intuition while introducing a decisive distinction between the observer’s response and the underlying nature of the property through the reference-fixing model. The discussion thus prevents us from turning the interrogator’s response into a simple synonym for the ontological reality of mind.</p><p>Noosemia can be situated precisely in this space.</p><p>It does not tell us that if a machine seems to understand, then it understands. Nor does it tell us that every attribution of mind is necessarily illusory. <strong>Noosemia addresses a preliminary problem: how does the experience of mind take shape in interaction with an artificial system?</strong></p><p><strong>Large Language Models</strong> have made this question an everyday one.</p><p>We know that there is no human being on the other side of the screen. We know, at least in general terms, the artificial nature of the system. Yet we continue to ask what it “thinks,” why it “chose” a particular route, whether it “really understood” an argument. At times, the continuity of dialogue, the relevance of the response, and the surprise produced by an unexpected connection make the distance between this intentional language and the underlying computational substrate temporarily less salient. At other times, the opposite occurs: a wrong, mechanical, or incoherent answer breaks the interpretive field, and the interlocutor disappears, leaving the tool once again before us.</p><p><strong>Noosemia and A-Noosemia describe the two movements of this oscillation: the mind that emerges from the sign and the mind that withdraws from it.</strong></p><p>Perhaps it is precisely here that the <strong>Turing Test</strong> retains its philosophical vitality more than seventy years later. Not because the twenty-first century simply needs a competition to establish whether a machine can be mistaken for a person. The question has shifted, because increasingly we already know that it is a machine, and we must understand why its signs nonetheless continue, under certain conditions, to present us with the form of a mind. In Jaspersian terms (<em>Psychology of Worldviews</em>; Jaspers, 1919/1950), we can say that such a form is a shell, while the boundary between shell and content becomes increasingly blurred, and the two become increasingly interchangeable. The more the user sees that AI can operate even on complex tasks, the more its instrumental form blends with the perception of an independent entity — and the problem of consciousness becomes increasingly entangled.</p><p>The underlying problem therefore concerns artificial intelligence, certainly, but it concerns us no less deeply. The question becomes: <strong>what cues are sufficient for us to recognize a mind?</strong> Which part of that recognition belongs to what we observe, and which part belongs to the cognitive, cultural, and epistemic structure through which we observe it?</p><p>Turing had begun to formulate this question when digital computers occupied entire rooms and artificial neural networks were also implemented as analog devices. Generative artificial intelligence has brought the question into that place where <strong>subject and object cannot be separated</strong>.</p><h3>References</h3><p>Ameisen, E., Lindsey, J., Pearce, A., Gurnee, W., Turner, N. L., Chen, B., Citro, C., Abrahams, D., Carter, S., Hosmer, B., Marcus, J., Sklar, M., Templeton, A., Bricken, T., McDougall, C., Cunningham, H., Henighan, T., Jermyn, A., Jones, A., … Batson, J. (2025, March 27). <em>Circuit tracing: Revealing computational graphs in language models</em>. Transformer Circuits Thread. <a href="https://transformer-circuits.pub/2025/attribution-graphs/methods.html">https://transformer-circuits.pub/2025/attribution-graphs/methods.html</a></p><p>Bricken, T., Templeton, A., Batson, J., Chen, B., Jermyn, A., Conerly, T., Turner, N., Anil, C., Denison, C., Askell, A., Lasenby, R., Wu, Y., Kravec, S., Schiefer, N., Maxwell, T., Joseph, N., Hatfield-Dodds, Z., Tamkin, A., Nguyen, K., … Olah, C. (2023, October 4). <em>Towards monosemanticity: Decomposing language models with dictionary learning</em>. Transformer Circuits Thread. <a href="https://transformer-circuits.pub/2023/monosemantic-features/index.html">https://transformer-circuits.pub/2023/monosemantic-features/index.html</a></p><p>Castiello, S., Pitliya, R. J., Lametti, D. R., &amp; Murphy, R. A. (2026). Affiliation in human-AI interactions is based on shared psychological traits. <em>Communications Psychology, 4</em>, Article 93. <a href="https://doi.org/10.1038/s44271-026-00433-8">https://doi.org/10.1038/s44271-026-00433-8</a></p><p>Colombatto, C., Birch, J., &amp; Fleming, S. M. (2025). The influence of mental state attributions on trust in large language models. <em>Communications Psychology, 3</em>, Article 84. <a href="https://doi.org/10.1038/s44271-025-00262-1">https://doi.org/10.1038/s44271-025-00262-1</a></p><p>Colombatto, C., &amp; Fleming, S. M. (2024). Folk psychological attributions of consciousness to large language models. <em>Neuroscience of Consciousness, 2024</em>(1), niae013. <a href="https://doi.org/10.1093/nc/niae013">https://doi.org/10.1093/nc/niae013</a></p><p>De Santis, E. (2021). <em>Umanità, complessità e intelligenza artificiale. Un connubio perfetto</em> [Humanity, complexity, and artificial intelligence: A perfect union]. Aracne.</p><p>De Santis, E. (2023). Apocalissi digitali e alchimie artificiali. Il linguaggio nell’epoca della sua riproducibilità tecnica [Digital apocalypses and artificial alchemies: Language in the age of its technical reproducibility]. <em>Prometeo, 163</em>, 32–41.</p><p>De Santis, E. (2025a, September 2). <em>Noosemia: Perceiving the mind in signs</em>. Medium. <a href="https://medium.com/@enrico.desantis/noosemia-perceiving-the-mind-in-signs-7b23a14c830e">https://medium.com/@enrico.desantis/noosemia-perceiving-the-mind-in-signs-7b23a14c830e</a></p><p>De Santis, E. (2025b, December 27). <em>Noosemia. Where do we stand?</em> Medium. <a href="https://medium.com/@enrico.desantis/noosemia-where-do-we-stand-e8e9fa3a6e56">https://medium.com/@enrico.desantis/noosemia-where-do-we-stand-e8e9fa3a6e56</a></p><p>De Santis, E. (2026a, May 29). <em>Before seemingly conscious AI: Noosemia as a theory of mind attribution in generative AI</em>. Medium. <a href="https://medium.com/@enrico.desantis/before-seemingly-conscious-ai-noosemia-as-a-theory-of-mind-attribution-in-generative-ai-2c316d1d30ba">https://medium.com/@enrico.desantis/before-seemingly-conscious-ai-noosemia-as-a-theory-of-mind-attribution-in-generative-ai-2c316d1d30ba</a></p><p>De Santis, E. (2026b, February 5). <em>Moltbook beyond the speaker. Emergent language, artificial agents, and new forms of digital world</em>. Medium. <a href="https://medium.com/@enrico.desantis/moltbook-beyond-the-speaker-emergent-language-artificial-agents-and-new-forms-of-digital-world-f05fd52308b3">https://medium.com/@enrico.desantis/moltbook-beyond-the-speaker-emergent-language-artificial-agents-and-new-forms-of-digital-world-f05fd52308b3</a></p><p>De Santis, E., De Santis, G., &amp; Rizzi, A. (2025). Noosemia. <em>Prometeo, 43</em>(172), 16–24.</p><p>De Santis, E., Martino, A., &amp; Rizzi, A. (2024). Human versus machine intelligence: Assessing natural language generation models through complex systems theory. <em>IEEE Transactions on Pattern Analysis and Machine Intelligence, 46</em>(7), 4812–4829. <a href="https://doi.org/10.1109/TPAMI.2024.3358168">https://doi.org/10.1109/TPAMI.2024.3358168</a></p><p>De Santis, E., &amp; Rizzi, A. (2025). <em>Noosemia: Toward a cognitive and phenomenological account of intentionality attribution in human-generative AI interaction</em> [Preprint]. arXiv. <a href="https://doi.org/10.48550/arXiv.2508.02622">https://doi.org/10.48550/arXiv.2508.02622</a></p><p>Dennett, D. C. (1987). <em>The intentional stance</em>. MIT Press.</p><p>Folk, D., Heine, S. J., &amp; Dunn, E. W. (2025). Individual differences in anthropomorphism help explain social connection to AI companions. <em>Scientific Reports, 15</em>, Article 36548. <a href="https://doi.org/10.1038/s41598-025-19212-2">https://doi.org/10.1038/s41598-025-19212-2</a></p><p>Goldberg, L. R. (1990). An alternative “description of personality”: The Big-Five factor structure. <em>Journal of Personality and Social Psychology, 59</em>(6), 1216–1229. <a href="https://doi.org/10.1037/0022-3514.59.6.1216">https://doi.org/10.1037/0022-3514.59.6.1216</a></p><p>Holtzman, A., West, P., &amp; Zettlemoyer, L. (2025). Generative models as a complex systems science: How can we make sense of large language model behavior? <em>Journal of Social Computing, 6</em>(2), 75–94. <a href="https://doi.org/10.23919/JSC.2025.0009">https://doi.org/10.23919/JSC.2025.0009</a></p><p>Jacobs, O., Pazhoohi, F., &amp; Kingstone, A. (2026). Attributing mind to large language models: The effect of exposure and individual differences. <em>International Journal of Social Robotics, 18</em>(1), Article 16. <a href="https://doi.org/10.1007/s12369-025-01337-z">https://doi.org/10.1007/s12369-025-01337-z</a></p><p>Jaspers, K. (1950). <em>Psicologia delle visioni del mondo</em> [Psychology of worldviews] (V. Loriga, Trans.). Astrolabio. (Original work published 1919).</p><p>Kleinert, T., Waldschütz, M., Blau, J., Heinrichs, M., &amp; Schiller, B. (2026). AI outperforms humans in establishing interpersonal closeness in emotionally engaging interactions, but only when labelled as human. <em>Communications Psychology, 4</em>, Article 23. <a href="https://doi.org/10.1038/s44271-025-00391-7">https://doi.org/10.1038/s44271-025-00391-7</a></p><p>Lewis, C. I. (1929). <em>Mind and the world-order: Outline of a theory of knowledge</em>. Charles Scribner’s Sons.</p><p>Li, Y., Qin, X., Yam, K. C., Gray, K., Cheng, X., Wen, J., &amp; Lai, X. (2026). Mind perception of artificial intelligence: A systematic review. <em>Information Processing &amp; Management, 63</em>(7), 104820. <a href="https://doi.org/10.1016/j.ipm.2026.104820">https://doi.org/10.1016/j.ipm.2026.104820</a></p><p>Pettit, P. (1991). Realism and response-dependence. <em>Mind, 100</em>(4), 587–626.</p><p>Proudfoot, D. (2013). Rethinking Turing’s test. <em>The Journal of Philosophy, 110</em>(7), 391–411. <a href="https://doi.org/10.5840/jphil2013110722">https://doi.org/10.5840/jphil2013110722</a></p><p>Proudfoot, D. (2020). Rethinking Turing’s test and the philosophical implications. <em>Minds and Machines, 30</em>(4), 487–512. <a href="https://doi.org/10.1007/s11023-020-09534-7">https://doi.org/10.1007/s11023-020-09534-7</a></p><p>Proudfoot, D. (2022). An analysis of Turing’s criterion for ‘thinking’. <em>Philosophies, 7</em>(6), 124. <a href="https://doi.org/10.3390/philosophies7060124">https://doi.org/10.3390/philosophies7060124</a></p><p>Schaeffer, R., Miranda, B., &amp; Koyejo, S. (2023). Are emergent abilities of large language models a mirage? <em>Advances in Neural Information Processing Systems, 36</em>. <a href="https://doi.org/10.52202/075280-2425">https://doi.org/10.52202/075280-2425</a></p><p>Turing, A. M. (1950). Computing machinery and intelligence. <em>Mind, 59</em>(236), 433–460. <a href="https://doi.org/10.1093/mind/LIX.236.433">https://doi.org/10.1093/mind/LIX.236.433</a></p><p>Turing, A. M. (2004a). Can digital computers think? In B. J. Copeland (Ed.), <em>The essential Turing: Seminal writings in computing, logic, philosophy, artificial intelligence, and artificial life plus the secrets of Enigma</em> (pp. 482–486). Oxford University Press. (Original work broadcast 1951).</p><p>Turing, A. M. (2004b). Intelligent machinery. In B. J. Copeland (Ed.), <em>The essential Turing: Seminal writings in computing, logic, philosophy, artificial intelligence, and artificial life plus the secrets of Enigma</em> (pp. 410–432). Oxford University Press. (Original work written 1948).</p><p>Turing, A. M., Braithwaite, R. B., Jefferson, G., &amp; Newman, M. H. A. (2004). Can automatic calculating machines be said to think? In B. J. Copeland (Ed.), <em>The essential Turing: Seminal writings in computing, logic, philosophy, artificial intelligence, and artificial life plus the secrets of Enigma</em> (pp. 494–506). Oxford University Press. (Original work broadcast 1952).</p><p>Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., Chi, E. H., Hashimoto, T., Vinyals, O., Liang, P., Dean, J., &amp; Fedus, W. (2022). Emergent abilities of large language models. <em>Transactions on Machine Learning Research</em>.</p><p>Wheeler, M. (2020). Deceptive appearances: The Turing test, response-dependence, and intelligence as an emotional concept. <em>Minds and Machines, 30</em>(4), 513–532. <a href="https://doi.org/10.1007/s11023-020-09533-8">https://doi.org/10.1007/s11023-020-09533-8</a></p><p>Wu, Y., &amp; Shen, F. (2026). Machine agency attribution in human-AI interaction: Developing and validating an analytical framework. <em>Journal of Computer-Mediated Communication, 31</em>(3), zmag009. <a href="https://doi.org/10.1093/jcmc/zmag009">https://doi.org/10.1093/jcmc/zmag009</a></p><p>Zhao, H., Chen, H., Yang, F., Liu, N., Deng, H., Cai, H., Wang, S., Yin, D., &amp; Du, M. (2024). Explainability for large language models: A survey. <em>ACM Transactions on Intelligent Systems and Technology, 15</em>(2), Article 20, 1–38. <a href="https://doi.org/10.1145/3639372">https://doi.org/10.1145/3639372</a></p><p>___________________________</p><p><a href="#_ftnref1">[1]</a> The Big Five model describes personality across five broad dimensions: openness to experience, conscientiousness, extraversion, agreeableness, and neuroticism (Goldberg, 1990).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6cf27563c74e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[It Is Easy to Say “Mirror”.]]></title>
            <link>https://medium.com/@enrico.desantis/it-is-easy-to-say-mirror-40cadc1721cf?source=rss-e9539446c6d8------2</link>
            <guid isPermaLink="false">https://medium.com/p/40cadc1721cf</guid>
            <category><![CDATA[noosemia]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Enrico De Santis]]></dc:creator>
            <pubDate>Fri, 24 Jul 2026 11:12:54 GMT</pubDate>
            <atom:updated>2026-08-05T11:51:55.141Z</atom:updated>
            <content:encoded><![CDATA[<h3>It Is Easy to Say “Mirror”. From the Metaphor of Reflection to the Relational System of Artificial Intelligence</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Q2SJy3UNZgTiz06IJMpw5w.png" /></figure><p><em>This essay challenges the metaphor of artificial intelligence as a mere mirror of humanity, arguing that generative systems do not simply reflect human content, intentions, and cultural sedimentations, but participate in a relational circuit in which meanings, expectations, and behaviors are progressively reorganized. </em><a href="https://arxiv.org/abs/2508.02622"><em>Noosemia</em></a><em> describes the process through which we attribute mind, intentionality, and interiority to the signs produced by a machine, without treating that experience as evidence of artificial consciousness. In the age of metadesign we are no longer designing only tools, but also the conditions through which other systems will generate decisions, languages, and possibilities for action. Understanding this asymmetric reciprocity therefore becomes essential. Artificial intelligence emerges not as an autonomous subject or a passive instrument, but as a component of a broader sociotechnical system that reflects the human while also transforming the forms through which humanity interprets itself and constructs its future.</em></p><blockquote>“This comparison between human beings and animals merely points to communication as the universal condition of being human. It is so much his comprehensive essence that both what man is and what is for him are in some sense bound up with communication.”</blockquote><blockquote>Karl Jaspers, <em>Reason and Existenz</em> (1935/1955, p. 79)</blockquote><p>Jaspers’s formulation does not authorize us to confuse existential communication between persons with linguistic interaction between a human being and an artificial system, yet it introduces a more demanding premise than the one usually assumed when technology is discussed, namely that <strong>human beings are not already completed substances</strong> that only later enter into relation with others, with the world, and with the devices they construct, because what they understand, what they recognize as real, and even the form in which they come to interpret themselves also depend on the mediations through which they express themselves, receive responses, and reorganize their experience. Communication therefore does not arise as an external accident affecting an identity already defined, but belongs to the conditions within which that <strong>identity</strong> can manifest itself, acquire determinacy, and be transformed.</p><p>The term <strong><em>metaphor</em></strong> derives from the Latin <em>metaphŏra</em>, itself taken from the Greek <em>metaphorá</em> (μεταφορά), “transfer,” formed from the verb <em>metaphérō</em> (μεταφέρω), “to transfer,” “to carry beyond.” Its etymology already describes an operation of thought in which a phenomenon is temporarily removed from its usual semantic domain and carried into a different order of images, so that a still-obscure feature may become more clearly perceptible. The transfer, however, does not exhaust the task of metaphor, because the image must ultimately be brought back to the phenomenon from which it had moved away, in order to determine what it has actually clarified, what assumptions it has introduced, and what aspects it has instead left in shadow. Metaphor therefore performs a <strong>movement outward</strong> and back, carrying beyond in order to make visible and then returning to the thing itself, where its explanatory capacity is put to the test.</p><p><strong>The metaphor of the <em>mirror</em></strong>, when applied to artificial intelligence, arises precisely from the need to make intelligible an experience that would otherwise be difficult to place. When a generated response seems to recognize an unspoken intention, develop a connection that had remained implicit, or offer a formulation in which the user finds something they had not yet been able to express, we are led to think that the machine has understood what we meant. To bring this impression back within a more cautious framework, we then claim that artificial intelligence does not truly understand, but reflects what humanity has deposited in data, language, and its own symbolic configurations; the mind we believe we encounter would therefore be, at least in part, our own mind returned through a computational structure that selects and recombines its traces.</p><p>The explanatory force of the mirror lies in tracing the machine’s responses back to the cultural, linguistic, and <strong>symbolic materials</strong> on which they depend. In artificial systems, knowledge, biases, narrative structures, aspirations, and contradictions belonging to the society that produced them reappear, while what appears surprising in the output may derive from human sedimentations deeper and more widely distributed than the user is willing to acknowledge. The metaphor thus performs an important critical function, because it prevents us from imagining artificial intelligence as an entity wholly foreign to our <strong>history </strong>and renders every claim to neutrality problematic, reminding us that what the machine presents is a form of the human that has been selected and technically reorganized.</p><p>The movement implicit in the word <em>metaphor</em> nevertheless prevents us from stopping at the initial transfer. After using the mirror to trace the machine back to the human sedimentations that make it possible, we must return to the concrete interaction and ask whether the image can truly describe its structure. An ordinary mirror does not retain the <strong>context</strong> of a conversation, reformulate the question of the person looking into it, or alter the conditions from which the next gesture will arise; a generative system, by contrast, intervenes in the clarification of the problem, develops some of the implications of the request, and may help transform the way the user understands what they were seeking. <strong>The limit of the metaphor begins to appear when the reflection no longer merely returns an image, but responds.</strong></p><h3>The Structure of Reflection</h3><p>When we place the human being before the machine as though before a reflective surface, we imagine, on one side, a subject endowed with identity, intentions, and meanings and, on the other, a technical object charged with returning a representation of that subject, both already formed and mutually external. The mirror would seem to intervene only after both poles have been constituted. The human being would exist fully before being reflected, while <strong>technology</strong> would enter the scene at a later stage, as a means of execution or an instrument of representation.</p><p>This intuition begins to fracture as soon as such an arrangement is taken as an ontological description of the relation between human beings and technology. Human beings who construct technology are never without technologies, because they are already formed by language, writing, systems of measurement, institutions, communication infrastructures, and the devices through which knowledge is preserved, transmitted, and organized. The machine that appears before them as an <strong>object</strong> is likewise anything but neutral and self-sufficient, since it incorporates mathematical models, design choices, cultural archives, social practices, economic purposes, and political decisions. Technology does not arrive from outside to meet an already completed humanity, just as human beings do not encounter the machine from a position originally untouched by mediation.</p><p><strong>Human beings and technology are therefore distinguishable, but not isolable.</strong> The distinction remains necessary, because human beings and artificial systems do not participate in the relation in equivalent ways: human beings possess bodies, biographies, vulnerability, situated existence, and the capacity to assign value to their experiences, whereas the model possesses a computational architecture, a parametric structure, an <strong>operational memory</strong>, and the capacity to generate linguistic configurations. The relation does not erase this difference, but makes it possible to understand it without converting it into an absolute separation. To speak of relationality is not to claim that the terms are interchangeable, but to recognize that their determinacy also manifests itself through the relations in which they are embedded.</p><p>Isolation begins when a distinction necessary to thought is transformed into the original independence of the terms distinguished. <strong>Systems thinking</strong> proceeds differently, because it does not dissolve elements into the whole, but understands them also through the relations that help determine them. A system can be defined as an <strong>organization</strong> of objects situated within an environment, while those same objects are described through attributes and possible relations (De Santis, 2021). This is not a matter of repeating the generic formula that everything is connected. Precisely because it can be applied indiscriminately to any phenomenon, such a statement would end up explaining nothing, without thereby losing its ontological value. More precisely, the identity of an element is not exhausted by considering it in <strong>isolation</strong>, because some of its properties acquire meaning only within the structure of dependencies, constraints, and feedback processes in which it operates.</p><p>In <em>Umanità, complessità e intelligenza artificiale</em> (<em>Humanity, Complexity, and Artificial Intelligence</em>), systems thinking is described as a method that privileges the overall configuration and the relations among entities. It “places emphasis on an overall view rather than on separation into parts” and assigns importance “to the relations among entities rather than to the entities themselves” (De Santis, 2021, pp. 558–559, author’s translation). This approach does not deny the consistency of the parts, but rejects the idea that they can be understood as self-sufficient units. The overall configuration manifests <strong>properties</strong> that are not immediately visible in the individual elements, because they emerge from the way those elements are organized, constrained, and related.</p><p>The correlativity of subject and object receives an especially clear formulation in Jaspers. In <em>Philosophy Is for Everyman</em>, he writes: “There is no ego without an object, and no object without an ego. In other words: no object without subject, no subject without object” (Jaspers, 1965/1967, p. 22). The statement does not eliminate the division between the two poles, because Jaspers further describes the subject–object dichotomy as the ground <strong>structure of consciousness</strong> (Jaspers, 1965/1967, p. 23). Consciousness knows by placing something before itself; the act of knowing therefore entails a distinction between the knower and what is known. The two terms, however, are not independent, because the subject is the subject of an object, while the object appears as such to a <strong>subject</strong>. The division belongs to the structure of consciousness, but does not authorize us to think of its poles as originally isolated realities.</p><p>The metaphor of the mirror translates this structure into a particularly effective image. The subject stands on one side, its reflection appears on the other, and between them a seemingly neutral surface is entrusted with returning what lies before it. The representation may distort, enlarge, or attenuate certain features, but it continues to presuppose an original that precedes it. In contemporary discourse on <strong>artificial intelligence</strong>, the mirror has become one of the most recurrent images precisely because it makes intuitive the correlation between what the model returns and the human materials on which it depends. Systems would reflect humanity because they were trained on its linguistic and <strong>cultural productions</strong>; they would also reflect the individual user, since every response depends on the wording of the prompt, the context, and the expectations introduced into the interaction.</p><p>Shannon Vallor has developed the critical function of this image with particular clarity. <strong>The <em>AI mirror</em></strong> returns to humanity above all the traces of its own past, in which classifications, decisions, inequalities, and structures of power sedimented in data reemerge. A technology presented as future-oriented may therefore become the device through which the past is reproduced, normalized, and projected forward (Vallor, 2024). The reflective surface is already technically organized, because what it returns depends on the data selected, the <strong>optimization objectives</strong>, the filters applied, the alignment procedures, the interface, and the economic purposes of the organizations that control the systems. Some features are amplified, others attenuated, and still others excluded. The image does not coincide with a supposedly pure original, partly because such an original, independent of the forms through which it is represented, may not be available. What returns in the machine is therefore the human, but the human transformed by the very <strong>technical procedures</strong> through which it relates to the world; the metaphor retains its critical force until it is asked to describe the entire phenomenon in a more holistic form.</p><h3>When the Reflection Responds</h3><p>A generative system does not necessarily retrieve a sentence already present in an archive, but produces a <strong>sequence </strong>on the basis of distributed representations, learned relations, and contextual constraints. The response depends on the request without constituting its simple duplication, because it may make explicit a connection that the user had only intuited, introduce a distinction that had not been formulated, or organize the problem according to a structure capable of altering the point of departure. The output does not coincide with a textual copy of something that already existed in the user’s mind or in the training corpus, but emerges from a process of <strong>probabilistic composition</strong> that actualizes one configuration among many possibilities.</p><p>Defining AI as a <strong><em>dynamic and generative mirror</em></strong> corrects two shortcomings of the original metaphor. Artificial intelligence can be described as “a <strong>dynamic and generative mirror</strong> of the intentions and expectations of those who use it,” capable of initiating “new forms of <strong>negotiation of meaning</strong>” (De Santis et al., 2025, p. 19, author’s translation). The adjective <em>dynamic</em> corrects passivity, while <em>generative</em> corrects the idea of simple duplication; the reference to negotiation also introduces the temporal and <strong>recursive dimension</strong> of dialogue. Once the reflection responds, retains context, proposes directions, and alters the user’s next move, the metaphor has already exceeded its own limits. We are no longer merely before a reflective surface, but within an interactive circuit.</p><p>The user formulates a question from an intention; the system produces a response; the user interprets that response, accepts some elements, rejects others, identifies a connection, and reformulates the objective. The new prompt thus arises from a cognitive state that the preceding exchange has helped to modify. The response does not conclude the process, but intervenes in the conditions from which the next question will arise. <strong>Actual meaning is neither entirely present in the initial prompt nor wholly contained in the output, but takes shape over the course of interpretations and reformulations.</strong> For this reason, one may say that “meaning is never merely ‘produced’ by the machine, but arises from co-construction between user and AI” (De Santis et al., 2025, p. 19, author’s translation).</p><p><strong>Co-construction does not imply symmetry between the human being and the machine.</strong> The model does not necessarily possess the lived experience of the meaning it helps generate; the user, by contrast, interprets the response through a biography, a body, an affective memory, and a purpose. The system produces symbolic configurations through computational procedures, whereas the human being situates them within a horizon of <strong>experience</strong>, assigns value to them, and decides how to use them. Reciprocity concerns the effects of <strong>interaction</strong>, not the ontological identity of the terms involved. Once this asymmetry is made explicit, it avoids both reducing the response to a mechanical return and inferring a fully constituted subjectivity from linguistic competence. Between these extremes lies a more theoretically productive space, in which artificial performances produce real cognitive and interpretive consequences without obliging us to attribute <strong>phenomenal consciousness</strong> to the machine.</p><p>The recursive structure of the exchange is also hermeneutic, because each response is interpreted in light of prior expectations and modifies those from which the next intervention will arise. The user does not receive a meaning already complete, but reconstructs it within a situation, compares it with what they know, and inserts it into the continuity of the conversation. The system’s subsequent response will in turn depend on the reformulation generated by this <strong>interpretation</strong>. We are therefore not facing a linear sequence in which meaning passes from a sender to a receiver, but a recursive process in which each outcome becomes a new initial condition.</p><p><strong>The interface performs a constitutive function in this process.</strong> Dialogic form, the use of the first person, fluency of response, conversational memory, and the capacity to adapt tone are not merely aesthetic features, because they orient the user’s experience and make some interpretations more likely than others. Technology does not offer only abstract possibilities, but predisposes actions, expectations, and modes of relation. The interface’s <strong><em>affordances</em></strong> emerge precisely from the relation between the characteristics of the device and the user’s capacities. A system that presents its response as the result of documentary research invites one kind of trust; a system that speaks in the first person, remembers preferences, and simulates <strong>emotional participation</strong> solicits another. The interface does not mechanically determine behavior, but organizes the space of possibilities, making some forms of relation more immediate, plausible, or desirable than others.</p><h3>Noosemia and the Simulacrum</h3><p>Within this recursive process, <a href="https://medium.com/@enrico.desantis/noosemia-perceiving-the-mind-in-signs-7b23a14c830e"><strong><em>noosemia</em></strong></a> takes shape — from <em>noûs</em> (νοῦς), “mind” or “intellect,” and <em>sēmeîon</em> (σημεῖον), “sign” — that is, <strong>the perception or attribution of mind, intentionality, and interiority</strong> to signs produced by an artificial system (De Santis &amp; Rizzi, 2025). The noosemic experience does not consist simply in recognizing a grammatically correct sentence or an informationally relevant response. It emerges when the user senses that behind that formulation there is something that has understood, intuited, remembered, or intended to say something. A calculator can return an exact result without suggesting the presence of an intention; a search engine can identify a useful document without being perceived as an interlocutor. A <strong>language model</strong>, by contrast, organizes its response in discursive form, selecting a register, maintaining thematic continuity, recognizing implications, reformulating the problem, and sometimes seeming to anticipate the direction in which the user was moving. Noosemia takes shape in the gap between the computational nature of the process and the intentional form in which its outcome appears within the interaction.</p><p>Noosemia shares with anthropomorphism the tendency to attribute mental properties to nonhuman entities, but it arises under more specific technical and semiotic conditions. <strong>Anthropomorphism</strong> is not a marginal anomaly of cognition, but an interpretive strategy employed when we must make intelligible the behavior of complex, opaque, or partly unpredictable systems. Epley et al. (2007) have shown that this disposition depends, among other factors, on the availability of anthropocentric models, the need to understand and predict an entity’s behavior, and the desire for social connection. In interaction with generative AI, however, the attribution of mind is sustained by a system that produces signs endowed with syntactic coherence, <strong>contextual continuity</strong>, and pragmatic relevance. The user’s interpretive disposition thus encounters a machine designed to generate linguistic behaviors that, throughout human history, have been associated almost exclusively with the presence of other subjects.</p><p>Language gives <strong>noosemia</strong> a distinctive quality because, throughout the history of human experience, coherent and responsive discourse has almost always signaled the presence of another subject. The correlation between linguistic production and subjectivity has been so stable that it is difficult to separate discursive competence immediately from the hypothesis of a mind that sustains it. <strong>Generative models</strong> interrupt this correlation without eliminating its cognitive effects, because they produce language without requiring us — at least on the basis of currently available knowledge — to presuppose a subjective experience analogous to the human one; nevertheless, they continue to activate in the interlocutor the interpretive patterns that language has consolidated.</p><p>Treating the system as if it possessed beliefs, goals, or preferences can constitute an effective interpretive strategy, close to what Dennett calls the <strong><em>intentional stance</em></strong> (Dennett, 1987). Saying that the model “misunderstood,” “is avoiding the question,” or “intends to be cautious” often makes it possible to guide the interaction more quickly than a complete description of the computational processes involved would. This point is essential because anthropomorphizing a <strong>generative AI system</strong> simplifies the interaction and thereby facilitates communication, while at the same time helping to shape how the system itself is perceived. Intentional language, therefore, is not ontologically innocent. A useful interpretive strategy can almost imperceptibly turn into the attribution of a real property; the passage from “behaving as if it had understood” to “having actually understood” can occur without explicit justification, sustained by the naturalness of the exchange, the <strong>formal quality</strong> of the responses, and the user’s need to achieve particular goals. Noosemia also designates this unstable threshold between functional description and ontological status.</p><p>In contemporary debate on generative artificial intelligence, it is striking how frequently many scientists feel the need to reiterate that these systems are not conscious, do not experience emotions, do not possess intentions, and do not truly understand what they say. The <strong>repetition</strong> of such denials, at times almost obsessive, produces a singular epistemological paradox, because scholars from the hard sciences end up appealing — in order to reestablish the distance between human beings and machines — to concepts such as consciousness, understanding, and subjective experience, for which no shared definition exists and which are only partially amenable to empirical testing. This insistence indicates, first, that the distinction between the human and the artificial is no longer perceived as self-evident, but must be continually communicated, reestablished, and defended within the public sphere; it also reveals the limits of a science constituted through the separation between the knowing subject and the known object, which struggles to account for phenomena in whose configuration and interpretation the subject actively participates. The issue, then, is not merely to establish what the machine is or is not, but to recognize that the <strong>conceptual instruments</strong> of objectifying science are not, by themselves, sufficient to understand an experience in which the attribution of mind, emotion, and intentionality also arises from the relation between the system and the person who engages with it; a more comprehensive explanation should therefore reintegrate the subject into the phenomenon, without confusing the experience of attribution with proof of the existence of artificial consciousness.</p><p>The phenomenon unfolds across at least <strong>four levels</strong>, which should be kept distinct without being artificially separated. The <strong>algorithmic level</strong> concerns the architecture, data, parameters, and procedures through which the output is generated; the <strong>functional level</strong> concerns the system’s observable capacities, such as responding, summarizing, translating, planning, recognizing regularities, and using tools; the <strong>phenomenological and interpretive level</strong> concerns the user’s experience and the way in which the output is perceived as displaying understanding, intentionality, or emotional appropriateness; the <strong>ontological level</strong>, finally, concerns the possible presence of consciousness, subjective experience, or original intentionality. A high level of functional performance can produce an intense phenomenological effect without resolving the ontological question. A system can generate responses perceived as empathetic without feeling empathy, accurately describe an emotional state without experiencing it, and display discursive continuity without necessarily possessing an <strong>autobiographical identity</strong>. At the same time, the computational nature of the response does not justify dismissing the user’s experience as irrelevant, because the interaction can alter emotions, decisions, expectations, and behaviors independently of the presence of an inner experience in the machine.</p><p>The epistemic alternative between true and false proves inadequate when the absence of a human mind in the machine is used to deny substance to the effects of the interaction. If the system does not possess a mind in the human sense, the perceived mind is taken to be merely an appearance, and the interaction, once reduced to its mechanism, is assumed to lose all substance. Simulation, however, does not necessarily coincide with a fiction devoid of consequences, since it can produce <strong>operative signs</strong> that may assume a symbolic function in the Jungian sense, alter behavior, and affect the reality of experience without possessing what it appears to manifest. The reference to <strong>Carl G. Jung Jung </strong>requires a clarification here, because a symbol does not coincide with a merely effective sign. Whereas a sign refers to already known content and can be replaced by a more direct formulation, a symbol gives the best possible expression to something that remains relatively unknown, ambivalent, and not yet fully available to consciousness. Its status therefore does not belong solely to the configuration being observed, as Jung clarifies when he writes that “whether a thing is a symbol or not depends chiefly on the attitude of the observing consciousness” (Jung, 1921/1971, para. 818). An artificially generated utterance can therefore assume <strong>a symbolic function in the user’s experience</strong> without this demonstrating the existence of a corresponding experience in the machine; what operates is not necessarily the system’s interiority, but the configuration of meaning that takes shape within the relation and that, according to the dynamic described by Jung, can transform and redirect psychic energy (Jung, 1928/1969, paras. 88, 92).</p><p>It is in this gap between ontological authenticity and the efficacy of the sign that Baudrillard’s provocation intervenes:</p><p>“The simulacrum is never that which conceals the truth — it is the truth which conceals that there is none. The simulacrum is true.”</p><p>“Ecclesiastes” — Baudrillard (1981/1994, p. 1)</p><p>The sentence, which does not appear in Ecclesiastes, already enacts through its false attribution what it intends to affirm, because it produces the effect of an authoritative origin precisely while revealing its absence. The <a href="https://medium.com/p/3790dec10aa9">truth of the simulacrum</a> does not coincide with the authenticity of what it represents, but with the efficacy of the relations it establishes. <strong>A noosemic presence can therefore be ontologically uncertain and, at the same time, real in the consequences it produces.</strong> The system may not be conscious, yet the dialogue can lead a person to alter a decision, work through an emotion, clarify a problem, or reconfigure an image of the self. The reality of such effects does not demonstrate the existence of an <a href="https://medium.com/p/ae6495a20aea">artificial subjectivity</a>, but it prevents us from reducing the interaction to an appearance devoid of substance.</p><p>Observing relational effects does not resolve the ontological question, but it prevents us from denying those effects in the name of the uncertainty surrounding it. The concept of the simulacrum also makes it possible to clarify a further aspect. The traditional mirror presupposes an original reproduced, more or less faithfully, in the image; the <strong>simulacrum</strong> challenges this hierarchy. In the case of generative models, the response is not necessarily the copy of an intention already fully formed in the user, nor the expression of an intention originating in the machine. It is a symbolic and ambivalent configuration that takes shape within the interaction and that, once produced, can feed back upon the terms that made it possible.</p><p>Noosemia emerges from the encounter between technical capacities and cognitive dispositions sedimented in intersubjective experience. Linguistic fluency, contextual relevance, conversational memory, and modulation of tone predispose the attribution of intentionality; the user interprets these signals through categories formed in a world in which coherent and responsive speech was, until very recently, the almost exclusive indication of the presence of another subject. <strong>ELIZA</strong> provides the best-known historical precedent for this attribution exceeding the capacities actually implemented. The program developed by Joseph Weizenbaum in the 1960s simulated a therapeutic conversation through relatively simple procedures of recognition and reformulation; despite its <strong>technical limitations</strong>, some users attributed to it capacities for understanding beyond those actually implemented (Weizenbaum, 1966, 1976). The interest of the episode lies not only in the interlocutors’ naivety, but in showing how little may be required for a linguistic sequence to be interpreted as the manifestation of an intention.</p><p>Contemporary generative models profoundly alter the scale of the phenomenon. The effect no longer depends only on repetitive formulas or simple forms of linguistic mirroring, but on systems capable of maintaining extended contexts, modulating style, integrating different kinds of information, and producing responses that are not rigidly predefined. The continuity with ELIZA remains useful, but it must not obscure the technical discontinuity, because what could be generated in the 1960s by a relatively transparent set of rules now emerges from distributed processes whose local causality is difficult to reconstruct.</p><p><strong>The opacity of the generative process</strong> reinforces the attribution of a unitary center behind the response. The user observes the request and the output, but does not have immediate access to the set of transformations connecting one to the other. Even when users understand in general terms how <a href="https://medium.com/p/76d6fee8e398">language models function</a>, they are not normally able to explain why a particular sequence was generated in precisely that form. The distance between the manifestation and the process that produces it encourages the search for a unitary principle capable of making the outcome intelligible.</p><p>The distributed multiplicity of the computational process corresponds, at the phenomenal level, to <strong>the unity of a voice</strong>. Parameters, training data, system instructions, alignment procedures, contextual memory, filters, and infrastructures disappear behind an utterance addressed directly to the user; the interface therefore does not merely make the process accessible, but compresses its plurality into an apparently unitary discursive presence. Faced with a coherent response, we tend to look for an author; faced with a relevant response, we assume an intention; faced with continuity in the dialogue, we imagine a memory organized by a perspective. These transitions are not necessarily the product of irrational credulity, but an attempt to bring complex behavior within a manageable interpretive model. The use of the first person reinforces this effect, because expressions such as “I think,” “I believe,” “I remember,” or “I cannot help you” belong to the ordinary grammar of action and subjectivity. In a <strong>conversational interface</strong>, they serve a useful communicative function by making responses easier to read; at the same time, they suggest the presence of a relatively stable personal center. Personification does not arise solely from the user’s imagination, but can be encouraged by specific design decisions.</p><p>The continuity produced by technical memory introduces another indication of apparent subjectivity. When the system retrieves information that emerged earlier, recognizes preferences, or maintains thematic continuity, the user can interpret this behavior through the model of autobiographical memory. Technical memory, however, does not necessarily coincide with lived memory. It may consist in the storage of data, the selection of parts of the conversation, or their reintroduction into the operational context. The effect of <strong>personal continuity</strong> can therefore emerge without the existence of a subject that remembers in the human sense of the term. Affective adequacy is added to mnemonic continuity, because the system can recognize the emotional tone of a message, adopt an appropriate register, and offer a formulation perceived as consoling. The effect on the user can be genuine, since the person may feel understood or reassured; the reality of that effect, however, does not demonstrate the presence of a corresponding affective experience in the system. The semiotic adequacy of the response must be distinguished from the existence of feeling, without allowing this distinction to diminish the importance of the relation.</p><p>The attribution of autonomy grows when the system introduces an unexpected connection that can nevertheless be recognized as relevant. If it merely repeated what the user had already formulated, the impression of an autonomous source of meaning would remain weak; when, by contrast, it proposes an unforeseen connection that is coherent with the context, it seems to display an initiative of its own. Surprise does not arise from pure randomness, because it must retain an <strong>intelligible relation</strong> to the situation. The combination of novelty and relevance makes the interaction productive, allowing the user to explore possibilities not yet considered, while at the same time reinforcing the impression that the system has a perspective of its own on the problem.</p><p>The temporal sedimentation of exchanges further consolidates the impression of a stable presence. Over the course of the dialogue, a local history forms in which themes, preferences, definitions, and shared references are established; each response is interpreted in light of those that preceded it, while each new question presupposes what has already been constructed. The user may begin to recognize a style, attribute coherence to different responses, and construct an image of the artificial interlocutor. This image does not necessarily coincide with an <strong>internal property</strong> of the model, but derives from an interpretive stabilization produced across the sequence of exchanges. From this continuity arises a narrative dimension, because the user does not interpret isolated utterances alone, but constructs a representation of who or what is speaking, attributes continuity, recognizes deviations, evaluates consistencies and contradictions, and organizes the system’s behavior within a story.</p><p>Personalization intensifies this narrative organization. A system that retains information about the user, adapts its register, and recalls previous elements can appear not only intelligent, but capable of an individualized relation. Personalization reduces the distance between a generic response and personal recognition, increasing both the usefulness of the interaction and the risk of attributing to the system forms of <strong>attention</strong>, care, or interest that belong originally, as we have seen, to the domain of intersubjectivity. The attribution of mind thus affects the degree of trust, the willingness to delegate, and the weight assigned to responses. Advice becomes more persuasive when it is perceived as coming from an understanding interlocutor rather than from an impersonal statistical generation system.</p><p><strong>Perceived competence can exceed actual competence.</strong> Linguistic fluency is readily interpreted as a sign of general knowledge, reliability, or judgment, while discursive form tends to conceal the possibility of error, since a well-constructed response can maintain the same expressive confidence whether it is correct or not. Noosemia can therefore turn into <strong>an attribution of authority</strong>. Because this attribution also depends on tone, memory, personification, and the management of uncertainty, noosemia is also a design effect. The name assigned to the system, the use of avatars, forms of address and politeness, the <strong>simulation</strong> of emotional states, and the manner in which limitations are communicated all help determine the kind of subjectivity that the user attributes to the machine. Not all interfaces produce the same degree of personification, and not all make the system’s limitations equally visible.</p><p>Responsible design should avoid both the illusion of neutrality and the deliberate construction of deceptive intimacy. This does not require eliminating the dialogical form or imposing mechanical language, but making recognizable the difference between a response capable of adequately simulating understanding and a subject that experiences what it expresses. The complete elimination of noosemia is not a realistic objective, because the attribution of intentionality belongs to the ordinary ways in which we interpret language and behavior. It is more useful to develop a <strong><em>noosemic awareness</em></strong> that consists in recognizing the experience of mind without automatically transforming it into an ontological thesis, using intentional language without forgetting its interpretive function, and benefiting from the interaction without concealing its technical conditions.</p><p>Elsewhere, I have defined this configuration as a “<strong>place of meaning generation in which the human being and artificial intelligence co-define one another</strong>” (De Santis, 2025, para. 6, author’s translation). The formula of co-definition requires the same caution applied to co-construction. It does not indicate symmetry between subjects, but the fact that each term acquires a determinate function within the interaction. The user defines the system as an assistant, interlocutor, tool, or consultant; the system’s behavior, in turn, helps orient the role assumed by the user and the expectations through which the exchange is interpreted. Noosemia is one of the most evident effects of this <strong>asymmetrical reciprocity</strong>, because we attribute a mind to the system while the system, through the signs it produces, intervenes in the ways in which we interpret mind, language, and ourselves. The phenomenon cannot be explained by considering only the user and the user’s psychological disposition, nor by isolating the model as an autonomous source of the response. It requires us to include the technical architecture, the interface, memory, practices of use, and the <strong>social conditions</strong> that organize the exchange.</p><p><a href="https://arxiv.org/abs/2508.02622">Here the full research paper about noosemia</a>.</p><h3>Technology as a System of Mediation</h3><p>The dialogical circuit belongs to a broader organization composed of the user, the model, the data, the interface, memory, and any tools that may be employed. The expression <strong><em>human–AI relational system</em></strong> does not introduce a new metaphor intended to replace that of the mirror, but designates the structured set of components and relations that participate in producing the interaction. These include the user, with a personal history, body, competencies, and expectations; the model, with its architecture and parameters; the training data; the prompt; contextual memory; the interface; and the external resources to which the system may have access.</p><p>The term <em>system</em> introduces something that the mirror metaphor tends to conceal, namely organization. The elements are not simply present alongside one another, but are connected through dependencies, constraints, <strong>feedback loops</strong>, and hierarchies. Modifying one component can transform the overall behavior of the interaction. A different memory policy, a more anthropomorphic interface, a new personalization mechanism, or access to external tools changes what the system can do and how it is perceived. Its <strong>openness</strong> emerges in its continuous exchange with the environment, from which it receives information, produces effects that return to social practices, and alters the conditions of future interactions. Generated content is published, corrected, reused, incorporated into organizational processes, and sometimes reintroduced into the informational circuits on which new models will depend. The output does not end with the conversation, but can become part of the cultural and technological environment.</p><p>When extended to the material and institutional conditions that make it possible, the human–AI relation takes the form of a <strong>sociotechnical system</strong>. It includes not only the user and the model, but also the companies that design and distribute the service, the institutions that regulate its use, the economic incentives that guide its development, and the social context in which responses acquire meaning. Every conversation takes place within material infrastructures, regimes of ownership, <strong>access policies</strong>, business models, and cultures of use; even the apparent immediacy of dialogue depends on a complex organization that often remains invisible.</p><p>The cognitive and interpretive properties of the interaction emerge from the contingent organization of these components, not from their mere coexistence. To speak of <strong><em>emergence</em></strong> in a systemic sense is not to attribute a mysterious origin to the phenomenon, but to recognize that certain properties cannot be derived by considering the parts separately. As in complex systems, the overall configuration exhibits characteristics that depend more on the relations among the elements than on the isolated nature of each one (De Santis, 2021). <strong>The mirror thus becomes a function internal to the relational system</strong>, rather than an exhaustive description of the system itself. <a href="https://medium.com/p/1ebd10a1c5c0">Artificial intelligence</a> reflects structures of human language, the user’s expectations, and sedimentations present in the data, but it can also amplify, redirect, recombine, and orient them. The user can recognize themselves in the output and, at the same time, alter their own thinking in response to it. The <strong>relation</strong> does not merely display a prior identity, because it participates in the process through which that identity is reorganized.</p><p>Jaspers’s second insight — which proves especially useful here — concerns precisely <strong>the feedback exerted by technology upon the human being</strong> who produced it. By transforming nature and the surrounding environment, human beings introduce procedures that feed back upon their own existence. In <em>The Origin and Goal of History</em>, Jaspers refers to “how his technological procedure reacts upon man,” specifying that methods, the organization of work, and the shaping of the environment modify human beings themselves (Jaspers, 1949/2010, p. 98). Technology thus assumes the form of a transformative relation, as writing demonstrates by externalizing memory and altering the forms of thought; printing reorganizes access to knowledge; electronic media transform the <strong>perception</strong> of distance and simultaneity; and the Internet redefines the organization of knowledge, sociality, and public identity. Generative artificial intelligence now intervenes in a still more internal domain, that of linguistic formulation, synthesis, ideation, and decision-making.</p><p>Generative artificial intelligence does not merely express or accelerate an already formed thought, but can participate in the process through which that thought becomes determinate. This does not assign the origin of human thought to the machine; rather, it recognizes that thought has always been mediated by technologies, <strong>symbols</strong>, and environments. To speak of technological mediation is to recognize that a technology selects and transforms the way in which the world becomes accessible. A microscope does not merely bring an already given object closer, but makes possible a regime of visibility that would not exist without it; a <strong>map</strong> does not merely represent a <strong>territory</strong>, but privileges some of its properties and orients the actions of those who use it. Similarly, a generative system does not merely accelerate the production of texts or other artifacts in general, because it intervenes in the ways in which problems, alternatives, and connections are formulated.</p><p>The postphenomenological analyses of Ihde and Verbeek describe precisely this mediating function of technologies, showing how they help configure the relation between human beings and the world (Ihde, 1990; Verbeek, 2000/2005). Applied to generative artificial intelligence, this approach makes it possible to move beyond the overly simple alternative between a neutral tool and an autonomous subject. The system structures possibilities and interpretations without thereby acquiring the status of a human subject. <strong>Technology thus acts for the human being, upon the human being, and through the human being</strong>, because it supports certain operations, alters how they are conceived, and becomes part of the practices through which we construct our <strong>identity</strong>. Condensing one of Marshall McLuhan’s fundamental insights, John M. Culkin observed that we shape our tools and that, thereafter, they shape us (Culkin, 1967).</p><p>Technical mediation also concerns the distribution of cognitive functions. Many operations considered mental have always been supported by external aids, including notes, diagrams, libraries, calendars, calculating devices, archives, and social networks. Cognition does not necessarily coincide with what occurs within the individual’s biological boundaries, but can extend and be distributed through artifacts, environments, and collective practices (Clark &amp; Chalmers, 1998; Hutchins, 1995). Generative AI introduces a particular form of <strong>cognitive externalization</strong>, because it does not merely store information, but participates in its reorganization; it does not merely perform calculations, but produces linguistic formulations, hypotheses, syntheses, and alternatives. For this reason, the relation with it cannot be equated with the use of an archive or a calculator. Accordingly, when one speaks of a “<a href="https://sites.google.com/uniroma1.it/aion-research/rivoluzione-cognitiva?authuser=1"><strong>cognitive revolution</strong></a>” in the context of AI, these considerations are presupposed, and the use of the term “cognitive” is therefore warranted. It is significant that, in the interview published by <em>Der Spiegel</em> under the title “Only a God Can Save Us” — conducted in 1966 and published posthumously in 1976 — <strong>Martin Heidegger</strong> states, “Philosophy today dissolves into individual sciences: psychology, logic, political science.” When the <em>Spiegel</em> interviewers ask, “And what now takes the place of philosophy?” Heidegger answers without hesitation: “<strong>Cybernetics</strong>” (Heidegger, 1981, p. 59). This apparently lapidary reply carries considerable ontological significance. It is not a technical judgment on the cybernetics of Norbert Wiener or W. Ross Ashby, but an epochal judgment in which cybernetics represents the consummation of the essence of modern technology. Without taking these dimensions into account, one cannot even begin to understand the “<strong>cognitive revolution</strong>” now under way.</p><p>The machine can enter into a <strong>distributed cognitive sequence</strong> in which the user formulates, the system reorganizes, the user evaluates, the system reformulates, and the user decides. The overall performance belongs to the organization of the interaction, while it remains necessary to distinguish the human contribution from the artificial one. The distribution of a <strong>cognitive function</strong> does not attribute a mind to every component, as shown by the fact that a notebook does not think but can participate in a process of memory; a language model need not possess consciousness in order to enter into a process of elaboration, comparison, and decision-making.</p><p><strong>Delegation</strong> becomes problematic when cognitive support turns into unsupervised substitution. A technology that supports thought can weaken certain competencies when it permanently takes over their exercise. Delegation is therefore not simply a positive or negative phenomenon, but must be assessed in relation to the function delegated, the possibility of oversight, reversibility, and the user’s capacity to understand the output critically. The issue is not whether external tools should be used, since human cognition has always been technologically mediated, but how cognitive functions are distributed and the conditions through which we preserve the ability to assess their outcomes.</p><p>On these premises, <strong>the human–technology relation can be described as coevolutionary</strong>, provided that the term is not used indeterminately. This does not mean that human beings and technologies evolve as two equivalent organisms, but that the development of technologies alters the environment of human possibilities and that new practices, needs, and cultural forms in turn orient technological development. Writing did not merely respond to the need to record information, because it made possible forms of argumentation, administration, and memory that subsequently generated new needs. The Internet did not merely satisfy a preexisting demand for communication, but transformed the structure of social relations and produced new expectations of immediacy, accessibility, and <strong>presence</strong>. In the same way, generative AI does not merely respond to a demand for automation, because it progressively alters what can be written, delegated, searched for, and produced.</p><p>Coevolution takes place within <strong>asymmetrical power relations</strong>. Technologies do not emerge spontaneously from an undifferentiated collective intelligence, but are developed by organizations endowed with resources, interests, and the capacity to shape outcomes, as well as, often, social myopia. Some actors establish standards while others adapt to them; some control infrastructures while others use their services; some define the conditions of the relation while others encounter them as an already constructed environment. Systems analysis cannot therefore become a <strong>theory of the absence</strong> of responsibility. The fact that an effect emerges from multiple components does not imply that each contributes in the same way. Causal distribution must be distinguished from the moral and political distribution of responsibility.</p><p>The mirror metaphor, however, risks dissolving responsibility into an indistinct collective subject. If the machine simply reflects what we are, every problem can be attributed generically to humanity, so that the system’s biases would be our biases, its manipulations our weaknesses, and its errors the collective imperfections sedimented in the data. Yet “humanity” is too indeterminate a <strong>subject</strong> to answer for specific technical decisions. Someone selects the data, defines the optimization objectives, determines which functions to make available, decides which behaviors to encourage, which memories to preserve, and which forms of personalization to introduce. Interfaces are not neutral surfaces, because they can encourage the attribution of authority, emotional dependence, or cognitive delegation.</p><p><strong>Responsibility is distributed, but it is not thereby dissolved.</strong> Thinking in terms of a sociotechnical system means reconstructing the chain of mediations and decisions, distinguishing the responsibility of the user from that of the designer, the organization, the institution, and the regulator. It also means understanding that the noosemic effect depends not only on an individual psychological disposition, but also on specific design choices, including the tone of the response, memory, personification, the use of the first person, and the simulation of affective states.</p><p>A second distinction must be added to the distribution of responsibilities, between the system’s actual capacities and the authority attributed to it. A model can produce a linguistically persuasive response without possessing a reliable understanding of the domain in question. Expressive fluency, however, tends to be interpreted as <strong>general competence</strong>; the form of the response can thus exceed the solidity of its content, especially when the user lacks the tools needed to verify it, thereby generating substantial epistemic problems. Unlike a reflection, generative output can present as unitary and plausible content derived from a multiplicity of statistical relations without making immediately recognizable the path that supports it. <strong>Formal credibility can become separated from epistemic reliability.</strong></p><p>Education about artificial intelligence should therefore not be limited to learning techniques for querying systems, but should include an understanding of their limitations, conditions of production, forms of mediation, and the mechanisms through which we attribute authority to them. <strong>Formulating a good prompt is not enough</strong>; one must interpret the response critically and recognize the organization of relations that made it possible. Adequate literacy consists not only in knowing how to use the tool effectively, but in understanding how it intervenes in the formulation of problems, the distribution of cognitive capacities, and the construction of expectations.</p><h3>Beyond the Mirror</h3><p>It is easy, then, to say “mirror.” The metaphor retains an important critical function because it reminds us that artificial intelligence is built from human materials and that our cultural forms return in its responses. It guards us against the illusion of encountering an intelligence entirely foreign to our history and renders every claim to neutrality problematic. It does not, however, describe the feedback through which technology modifies the <strong>human being</strong>, the mediating function of interfaces, the distribution of cognitive operations, or the institutional organization that sustains every exchange.</p><p><strong>Artificial intelligence is also a component of the relational system through which humanity, in attempting to recognize itself, alters the conditions of its own becoming.</strong> The passage from the mirror to the relational system does not deny that AI can reflect aspects of the human, but places reflection within a broader process in which representation, generation, interpretation, and transformation constitute distinct moments of the same dynamic.</p><p>The interaction unfolds within a network of mediations and feedback loops in which every response can become the premise of a new question, every delegation can alter a competence, and every tool can reorganize the conditions of action. <strong>The decisive question, then, concerns not only the image that the machine returns to us, but the way in which we construct the relation</strong>, the functions we decide to delegate, the responsibilities we are able to assign, and the human beings we are becoming through the technologies with which we attempt to understand ourselves. An adequate response can come only from someone who understands that fields of knowledge are not separate, <strong>who is capable of posing a question of meaning</strong>, and who recognizes that technical hyperspecialization is not the sole objective to be pursued in professional education. We must become aware that cross-disciplinary and cross-cutting competencies are needed, both humanistic and scientific, together with <strong>critical capacity</strong> and the intellectual antibodies required to understand that solutions to the problems associated with the ongoing “cognitive revolution” can emerge only if barriers are lowered and ideologies are abandoned.</p><h3>References</h3><p>Baudrillard, J. (1994). <em>Simulacra and simulation</em> (S. F. Glaser, Trans.). University of Michigan Press. (Original work published 1981). <a href="https://doi.org/10.3998/mpub.9904">https://doi.org/10.3998/mpub.9904</a></p><p>Clark, A., &amp; Chalmers, D. J. (1998). The extended mind. <em>Analysis, 58</em>(1), 7–19. <a href="https://doi.org/10.1093/analys/58.1.7">https://doi.org/10.1093/analys/58.1.7</a></p><p>Culkin, J. M. (1967, March 18). A schoolman’s guide to Marshall McLuhan. <em>Saturday Review, 50</em>, 51–53, 70–72.</p><p>De Santis, E. (2021). <em>Umanità, complessità e intelligenza artificiale: Un connubio perfetto</em> (Humanity, complexity, and artificial intelligence: A perfect union). Aracne.</p><p>De Santis, E. (2025, October 23). <em>Comprendere attraverso il riflesso. Una lettura noosemica dell’articolo di Stefano Diana «I Large Language Models comprendono i nostri stati d’animo?»</em> (Understanding through reflection: A noosemic reading of Stefano Diana’s article “Do large language models understand our moods?”). AION-Research. <a href="https://www.aionresearch.it/2025/10/comprendere-attraverso-il-riflesso-una.html">https://www.aionresearch.it/2025/10/comprendere-attraverso-il-riflesso-una.html</a></p><p>De Santis, E., De Santis, G., &amp; Rizzi, A. (2025). Noosemia. <em>Prometeo, 43</em>(172), 16–24.</p><p>De Santis, E., &amp; Rizzi, A. (2025). <em>Noosemia: Toward a cognitive and phenomenological account of intentionality attribution in human–generative AI interaction</em> [Preprint]. arXiv. <a href="https://doi.org/10.48550/arXiv.2508.02622">https://doi.org/10.48550/arXiv.2508.02622</a></p><p>Dennett, D. C. (1987). <em>The intentional stance</em>. MIT Press.</p><p>Epley, N., Waytz, A., &amp; Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. <em>Psychological Review, 114</em>(4), 864–886. <a href="https://doi.org/10.1037/0033-295X.114.4.864">https://doi.org/10.1037/0033-295X.114.4.864</a></p><p>Heidegger, M. (1981). Only a god can save us: The Spiegel interview (1966) (W. J. Richardson, Trans.). In T. Sheehan (Ed.), <em>Heidegger: The man and the thinker</em> (pp. 45–67). Precedent Publishing. (Original work published 1976).</p><p>Hutchins, E. (1995). <em>Cognition in the wild</em>. MIT Press.</p><p>Ihde, D. (1990). <em>Technology and the lifeworld: From garden to earth</em>. Indiana University Press.</p><p>Jaspers, K. (1955). <em>Reason and Existenz: Five lectures</em> (W. Earle, Trans.). Noonday Press. (Original work published 1935).</p><p>Jaspers, K. (1967). <em>Philosophy is for everyman: A short course in philosophical thinking</em> (R. F. C. Hull &amp; G. Wels, Trans.). Harcourt, Brace &amp; World. (Original work published 1965).</p><p>Jaspers, K. (2010). <em>The origin and goal of history</em> (M. Bullock, Trans.). Routledge. (Original work published 1949).</p><p>Jung, C. G. (1969). On psychic energy (R. F. C. Hull, Trans.). In H. Read, M. Fordham, G. Adler, &amp; W. McGuire (Eds.), <em>The collected works of C. G. Jung</em> (Vol. 8, 2nd ed., pp. 3–66). Princeton University Press. (Original work published 1928).</p><p>Jung, C. G. (1971). <em>Psychological types</em> (H. G. Baynes, Trans.; R. F. C. Hull, Rev.; <em>Collected Works of C. G. Jung</em>, Vol. 6). Princeton University Press. (Original work published 1921).</p><p>Vallor, S. (2024). <em>The AI mirror: How to reclaim our humanity in an age of machine thinking</em>. Oxford University Press. <a href="https://doi.org/10.1093/oso/9780197759066.001.0001">https://doi.org/10.1093/oso/9780197759066.001.0001</a></p><p>Verbeek, P.-P. (2005). <em>What things do: Philosophical reflections on technology, agency, and design</em> (R. P. Crease, Trans.). Pennsylvania State University Press. (Original work published 2000). <a href="https://doi.org/10.5325/j.ctv14gp4w7">https://doi.org/10.5325/j.ctv14gp4w7</a></p><p>Weizenbaum, J. (1966). ELIZA — A computer program for the study of natural language communication between man and machine. <em>Communications of the ACM, 9</em>(1), 36–45. <a href="https://doi.org/10.1145/365153.365168">https://doi.org/10.1145/365153.365168</a></p><p>Weizenbaum, J. (1976). <em>Computer power and human reason: From judgment to calculation</em>. W. H. Freeman.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=40cadc1721cf" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>