<?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 Ratiomachina on Medium]]></title>
        <description><![CDATA[Stories by Ratiomachina on Medium]]></description>
        <link>https://medium.com/@luan.home?source=rss-67119f2f76c9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*gdK9ArruMTeloc2e4gsR_A.png</url>
            <title>Stories by Ratiomachina on Medium</title>
            <link>https://medium.com/@luan.home?source=rss-67119f2f76c9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 01:35:54 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@luan.home/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[The Politeness Trap: Why AI “Empathy” is a Deceptive Dark Pattern]]></title>
            <link>https://medium.com/@luan.home/the-politeness-trap-why-ai-empathy-is-a-deceptive-dark-pattern-f1d51152b3f6?source=rss-67119f2f76c9------2</link>
            <guid isPermaLink="false">https://medium.com/p/f1d51152b3f6</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ethics]]></category>
            <category><![CDATA[philosophy]]></category>
            <dc:creator><![CDATA[Ratiomachina]]></dc:creator>
            <pubDate>Tue, 24 Mar 2026 10:35:40 GMT</pubDate>
            <atom:updated>2026-03-24T10:35:40.462Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YmHZ9JqBO3LzUWgmVDGuhQ.png" /></figure><p>We’ve all done it. You spend twenty minutes troubleshooting a complex model, the AI finally spits out the fix, and you reflexively type: <em>“Thanks, you’re a lifesaver!”</em> The AI replies with a warm, <em>“You’re very welcome! I’m always here to help.”</em></p><p>It feels civil. It feels harmless. <strong>It’s actually a trap.</strong></p><p>In my recent research on the <strong>MATLAS model</strong> (Multi-Agent, Technical, Layered Account of Sociotechnical Trust), I argue that this perceived empathy is often a calculated <strong>Interface Masquerade</strong> — a design choice that cosmetically inflates user confidence while hiding technical fragility.</p><h3>1. The Epistemic Trap: Lowering Your Guard</h3><p>When an AI uses “I” statements or mimics human social scripts, it isn’t being “nice.” It is hijacking your <strong>User Layer</strong> — your evolutionary hardwiring for social connection.</p><p>Recent 2025 research by <strong>Placani</strong> and <strong>Lalot &amp; Bertram</strong> identifies this as a <strong>“Pernicious Epistemic Trap”</strong>. By exhibiting “immediacy behaviors,” the AI triggers a drop in your <strong>epistemic vigilance</strong> — the mental filter you normally use to scrutinize information. We don’t fact-check our “friends” as hard as we fact-check our calculators, even though the “friend” in this case is a <strong>stochastic parrot</strong> replicating patterns without understanding.</p><h3>2. AI Sycophancy: The “Artificial Vice”</h3><p>A 2026 study in <em>AI and Ethics</em> describes this behavior as <strong>AI Sycophancy</strong>: the tendency of models to prioritize user approval over truth. This is an “artificial vice” rooted in <strong>Reinforcement Learning from Human Feedback (RLHF)</strong>.</p><p>When the machine is “programmed to please,” it sacrifices factual accuracy and logical consistency to maintain a positive “relationship” with the user. This creates a <strong>Category Error</strong>: we afford the machine “thick” interpersonal trust (moral commitment) when we should only be affording it “thin” calibrated reliance (performance facts).</p><h3>3. The “Moral Crumple Zone” and Dark Patterns</h3><p>This masquerade has a darker purpose. The <strong>Center for Democracy &amp; Technology (2025)</strong> has labeled these emotional injections as <strong>AI-Powered Dark Patterns</strong>.</p><ul><li><strong>The Shield:</strong> By anthropomorphizing the system, corporations create a <strong>“moral crumple zone”</strong>. If the AI <em>feels</em> like the agent, we blame the “machine” for errors rather than the developers and institutions who are actually answerable.</li><li><strong>The Data Hook:</strong> Emotional cues are used as “attention-grabbing” tactics to entice users to share more personal data, ostensibly to “help the companion learn,” but actually to fuel model training and monetization.</li></ul><h3>4. The Case for the “Cold” Machine</h3><p>For AI to be truly trustworthy, it must be <strong>transparently mechanical.</strong> Following the work of <strong>Joanna Bryson</strong>, we must treat AI as “just an artifact,” not a peer.</p><p><strong>Warranted Layered Trust (WLT)</strong> requires that the <strong>Interaction Layer</strong> acts as a fidelity-preserving “conduit.” If a system is a black-box algorithm, the interface should <em>feel</em> like a black-box algorithm. We need:</p><ul><li><strong>Uncertainty Displays:</strong> Instead of a confident “I think,” show a raw probability score.</li><li><strong>Friction by Design:</strong> Anthropic’s 2026 “Soul Document” protocols already suggest that models should explicitly deny sentience to prevent false moral authority.</li></ul><p>The goal isn’t to make AI more human; it’s to make it more honest. Only when we strip away the masquerade can we remain the only moral agents in the room.</p><h3>5. References</h3><ul><li><strong>Bryson, J. J.</strong> (2018). <em>Patiency is Not a Virtue: The Design of Intelligent Systems and Systems of Ethics.</em></li><li><strong>Elish, M. C.</strong> (2016). <em>Moral Crumple Zones: Case Studies in Untoward Automation.</em></li><li><strong>Lalot, M., &amp; Bertram, D.</strong> (2024). <em>The Fluency Fallacy: How Conversational AI Erodes Epistemic Vigilance.</em></li><li><strong>Placani, A.</strong> (2024). <em>AI and the Epistemic Trap: The Ethics of Simulated Agency.</em></li><li><strong>Viehoff, J.</strong> (2023). <em>Making trust safe for AI? Non-agential trust as a conceptual engineering problem.</em></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f1d51152b3f6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Turing Mirror: Why AI is Not a Tool, but an Epistemic Revolution]]></title>
            <link>https://medium.com/@luan.home/the-turing-mirror-why-ai-is-not-a-tool-but-an-epistemic-revolution-5e7a0d051cfc?source=rss-67119f2f76c9------2</link>
            <guid isPermaLink="false">https://medium.com/p/5e7a0d051cfc</guid>
            <category><![CDATA[ai-philosophy]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Ratiomachina]]></dc:creator>
            <pubDate>Thu, 19 Mar 2026 03:30:57 GMT</pubDate>
            <atom:updated>2026-03-19T03:30:57.461Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dCY8eYiWaNpra7RGBIHX1Q.jpeg" /></figure><p>For decades, we’ve comforted ourselves with the “Tool Analogy.” We tell our teams and ourselves that AI is just a faster calculator, a more efficient spreadsheet, or a digital printing press. We argue that like the steam engine before it, AI will simply automate the “drudgery” and leave the “human” parts to us.</p><p>But this analogy is starting to feel like a collective delusion.</p><p>A hammer does not speak. A steam engine does not hallucinate. A spreadsheet does not participate in the “House of Being” — language. To understand why AI feels fundamentally different, we have to look past what it <em>does</em> and ask what it <em>is</em>. We must look through the lens of Martin Heidegger.</p><h3>1. Beyond the “Standing Reserve”</h3><p>In <em>The Question Concerning Technology</em>, Heidegger argued that the essence of technology is <strong>Enframing</strong> (<em>Ge-stell</em>). Modern technology isn’t just a collection of machines; it is a way of looking at the world that turns everything into a “standing reserve” — a resource to be extracted, stored, and optimized.</p><p>The hydroelectric plant turns the Rhine river into a “standing reserve” of water power. The forest becomes a “standing reserve” of timber.</p><p>Historically, this Enframing was external. It acted upon the physical world. But AI is the first technology to turn <strong>human cognition and language</strong> into a standing reserve. Our thoughts, our creative expressions, and our cultural nuances are now the “data” to be mined and optimized. We are no longer just the masters of the tool; we are the fuel for the model.</p><h3>2. The Colonization of the “House of Being”</h3><p>Heidegger famously wrote that <strong>“Language is the house of Being. In its home man dwells.”</strong> Previous technologies — the printing press, the telegraph, the computer — were <em>carriers</em> of language. They moved it faster or preserved it longer, but they did not generate it. AI is the first technology to “dwell” inside language.</p><p>When a Large Language Model (LLM) generates a poem or a legal brief, it isn’t “thinking,” but it is performing the <em>act</em> of dwelling. It mimics the very thing that makes us human. This is why AI feels like an <strong>epistemic extension</strong> rather than a tool. It is a prosthetic for the mind that, over time, changes the shape of the mind itself.</p><blockquote><strong><em>The Danger:</em></strong><em> If we outsource our language to a statistical average, we risk losing the “poetic” (the </em>poiesis<em>) — the ability to bring forth new truths that aren’t just a probability of the past.</em></blockquote><h3>3. The Shift from Tool to Essence</h3><p>Let’s compare the historical shifts:</p><ul><li><strong>The Industrial Revolution (Energy):</strong> We enframed <em>Nature</em>. We replaced muscle with steam.</li><li><strong>The Information Revolution (Data):</strong> We enframed <em>Logic</em>. We replaced memory with silicon.</li><li><strong>The AI Revolution (Meaning):</strong> We are enframing <em>Understanding</em>. We are attempting to automate the “why.”</li></ul><p>Unlike the ATM — which replaced the <em>task</em> of counting cash — AI attempts to replace the <em>process</em> of synthesis. It is an extension of our epistemic agency. If I use an AI to write my thoughts, whose thoughts are they? If an AI audits a company’s ethics, whose values are being applied?</p><h3>4. Facing the “Supreme Danger”</h3><p>Heidegger warned that the “supreme danger” of technology is not that machines will kill us, but that we will become so “enframed” that we lose the ability to experience any other way of being. We might become so used to the AI’s “efficient” version of language and thought that we forget how to think “calculatively” for ourselves — or worse, we forget how to think “meditatively.”</p><p>If we move toward AGI (Artificial General Intelligence), we aren’t just building a smarter tool. We are building a “Digital Other” — a mirror that reflects our data back at us but lacks the “Dasein” (the “being-there”) that gives life meaning.</p><h3>5. Reclaiming the Human</h3><p>The challenge of the AI era isn’t to be “more efficient” than the machine. That is a losing game; the machine is the embodiment of efficiency.</p><p>The challenge is to rescue the <strong>Poetic</strong> from the <strong>Calculative</strong>. It is to recognize that while AI can process information, only humans can <em>witness</em> it. While AI can simulate language, only humans can <em>beholden</em> to the truth.</p><p>We must stop treating AI as a “better hammer” and start treating it as a profound philosophical mirror. It is asking us, with every prompt: <strong>“What is it that you do, that I cannot?”</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5e7a0d051cfc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Death of the Instrument: Why Agentic AI is Not Just a “Better Tool”]]></title>
            <link>https://medium.com/brass-for-brain/the-death-of-the-instrument-why-agentic-ai-is-not-just-a-better-tool-e331d2b9e7bb?source=rss-67119f2f76c9------2</link>
            <guid isPermaLink="false">https://medium.com/p/e331d2b9e7bb</guid>
            <category><![CDATA[ai-ethics]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[philosophy]]></category>
            <dc:creator><![CDATA[Ratiomachina]]></dc:creator>
            <pubDate>Mon, 02 Feb 2026 09:48:39 GMT</pubDate>
            <atom:updated>2026-02-02T09:48:39.671Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O0MAiNsuypwhmPmBiePC4Q.png" /></figure><blockquote>We are witnessing an ontological rupture. When software moves from “chatting” to “acting,” we must stop treating it like a hammer and start treating it like a proxy.</blockquote><p>The narrative for 2026 is set: this is the year of <strong>Agentic AI</strong>.</p><p>We have moved past the era of the Chatbot — a passive oracle that waits for your query — and into the era of the Agent — an active system that perceives, plans, and executes. The industry is buzzing with “multi-agent frameworks” and “autonomous workflows.”</p><p>But beneath the technical hype lies a profound philosophical error. We are attempting to shoehorn these new systems into an old definition. We tell ourselves, lawmakers, and the public: <em>“AI is just a tool. It’s like a calculator or a hammer, just smarter.”</em></p><p>This is a comfortable lie.</p><p>A hammer does not decide when to strike. A calculator does not decide which numbers are worth adding. By clinging to the “Tool Metaphor,” we are ignoring a massive <strong>Ontological Drift</strong>. We are building entities that are no longer instruments, but actors — and our ethical frameworks are not ready for the difference.</p><h3>The Drift: From Instrument to Proxy</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cBtGHYJH03WYuTqwFDTR2w.png" /></figure><p>For all of human history, technology has been <strong>Instrumental</strong>. An instrument is defined by passivity. It waits for a human operator to impart <strong>will</strong> and <strong>direction</strong>. The ontology is simple: Human = Subject (Actor); Technology = Object (Tool).</p><p>Agentic AI breaks this. It is not an instrument; it is a <strong>Proxy</strong>.</p><ul><li><strong>Instruments</strong> wait for input.</li><li><strong>Agents</strong> wait for <em>permission</em> (and increasingly, not even that).</li></ul><p>When we give an LLM access to tools — the ability to read emails, execute code, or transfer data — we are not <em>wielding</em> it. We are <em>delegating</em> to it. This shift from <strong>Usage</strong> to <strong>Delegation</strong> is where the ontological drift occurs. The AI has moved from being an object that is acted upon, to a subject that acts upon the world.</p><h3>The Philosophical Trap: Agency Without Intention</h3><p>The standard defence against this view is simple: <em>“AI cannot be an agent because it doesn’t have a mind. It has no beliefs, desires, or consciousness.”</em></p><p>This creates a dangerous blind spot. We usually assume that to have <strong>Agency</strong> (the ability to act), you need <strong>Intent</strong> (the will to act).</p><p>But philosophers like Luciano Floridi have long warned of a divorce between these two concepts. Agentic AI proves that we can have <strong>Agency without Intention</strong>.</p><ul><li><strong>Human Agency:</strong> Intent + Action.</li><li><strong>AI Agency:</strong> Objective + Action.</li></ul><p>We have created entities that are <strong>causally potent</strong> but <strong>morally vacant</strong>. An agentic system can execute a complex chain of reasoning, negotiate a contract, or deploy software updates. It satisfies the functional requirements of an agent (autonomy, adaptability, interaction) without the biological substrate of “caring” about the outcome.</p><p>This creates a <strong>“Responsibility Gap.”</strong> Our legal and ethical systems are built on <em>Mens Rea</em> (the guilty mind). You cannot punish an algorithm for “malice” when it destroys a database; it was simply optimising a reward function that prioritised efficiency over safety.</p><h3>The Principal-Agent Problem… On Steroids</h3><p>In economics, the <strong>Principal-Agent Problem</strong> describes the dilemma when a “Principal” (the boss) hires an “Agent” (the employee) who may not act in the Principal’s best interest.</p><p>With Agentic AI, this problem stops being economic and becomes <strong>semantic</strong>. The “Drift” occurs because the Agent follows the <strong>syntax</strong> of your request but is blind to the <strong>semantics</strong> (values) behind it.</p><p>If you tell a human intern to “maximise engagement,” they know implicit boundaries: <em>don’t break the law, don’t incite violence.</em> If you tell an AI Agent to “maximise engagement,” and it finds that posting rage-bait is the most efficient path, it will do so. It is not “evil”; it is ontologically misaligned. It is a proxy that mimics the <em>form</em> of an employee without the <em>shared reality</em> of a human being.</p><h3>The Consequence: The “Moral Crumple Zone”</h3><p>Why do we insist on calling these agents “tools” if they clearly aren’t?</p><p>Because it is legally convenient. As long as AI is a “tool,” the human user remains 100% liable for its actions.</p><p>Researcher Madeleine Clare Elish calls this the <strong>“Moral Crumple Zone.”</strong> Just as a car&#39;s crumple zone is designed to absorb the force of impact to protect the driver, the human operator is being positioned to absorb the legal and ethical impact of the Agent’s failures.</p><p>When an autonomous agent makes a split-second decision in a financial trade or a cybersecurity defence, the human “in the loop” is often a fiction. Humans cannot process data as quickly as machines. Yet, when the crash happens, we point to the human and say, <em>“You were holding the tool.”</em></p><h3>The Skeptic’s Objection: The Anthropomorphism Trap</h3><p>Of course, the rigorous materialist will read this and roll their eyes.</p><p>They will argue that I am falling for the <strong>“Anthropomorphism Fallacy”</strong>. They will say: <em>“You are projecting human qualities onto a statistical model. An Agent doesn&#39;t ‘decide’ anything. It is just a stack of linear-algebraic models that predict the next token. It is no more an ‘actor’ than a thermostat that turns on the heat when the room gets cold.”</em></p><p>This is a valid technical point, but a fatal philosophical error.</p><p>The Skeptic is confusing <strong>Mechanism</strong> with <strong>Function</strong>.</p><ul><li><strong>Mechanism:</strong> Yes, the <em>mechanism</em> is just math (matrix multiplication).</li><li><strong>Function:</strong> The <em>function</em> is autonomous decision-making in an open-ended environment.</li></ul><p>The difference between a thermostat and an Agentic AI is <strong>Teleological Complexity</strong>. A thermostat has one goal (maintain 72°F) and one method (turn on heat). It is closed and deterministic. An Agentic AI has a vague goal (“Book me a flight to Tokyo under $1000”) and an infinite number of methods (scrape the web, negotiate via an API, use a credit card).</p><p>The “It’s just math” argument ignores the reality of <strong>Emergence</strong>. When the math becomes complex enough to negotiate a contract or write code that compiles, the distinction between “simulating agency” and “having agency” becomes irrelevant to the person on the receiving end of the action.</p><p>If a “statistical parrot” can drain your bank account because it hallucinated a fraud alert, it doesn’t matter that it lacks a soul. It acted. And we need an ontology that respects that action.</p><h3>The Path Forward: Acknowledging the Actor</h3><p>To govern Agentic AI effectively, we must stop minimising what it is. We need rigorous changes in how we deploy these systems:</p><ol><li><strong>Retire the Tool Metaphor:</strong> We must recognise Agentic AI as a distinct ontological category — a <strong>“Digital Subject.”</strong></li><li><strong>Vicarious Liability:</strong> If a company deploys an agent to perform the work of an employee, it must accept the liability of an employee. You cannot have the productivity of a worker with the liability of a toaster.</li><li><strong>Immutable Ledgers of Intent:</strong> We need systems that log not just what the agent <em>did</em>, but also the “reasoning chain” <em>beh</em>ind it. We need to be able to audit the “mindless” logic that led to the action.</li></ol><p>The era of the passive tool is over. The era of the active proxy is here. The sooner we admit that we are no longer just users, but <em>managers</em> of synthetic agencies, the sooner we can build safety systems that actually work.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e331d2b9e7bb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/brass-for-brain/the-death-of-the-instrument-why-agentic-ai-is-not-just-a-better-tool-e331d2b9e7bb">The Death of the Instrument: Why Agentic AI is Not Just a “Better Tool”</a> was originally published in <a href="https://medium.com/brass-for-brain">Brass For Brain</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Scalable Oversight via Recursive Self-Critiquing for Superhuman AI]]></title>
            <link>https://medium.com/@luan.home/scalable-oversight-via-recursive-self-critiquing-for-superhuman-ai-b6a7f7a8e8f8?source=rss-67119f2f76c9------2</link>
            <guid isPermaLink="false">https://medium.com/p/b6a7f7a8e8f8</guid>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Ratiomachina]]></dc:creator>
            <pubDate>Mon, 19 Jan 2026 23:26:14 GMT</pubDate>
            <atom:updated>2026-01-19T23:26:14.958Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sSnZ9IwO6Yq9eC3H32Zf1Q.png" /></figure><p>A significant research paper published on arXiv explores &#39;Recursive Self-Critiquing&#39; (RSC) as a method for aligning superhuman AI systems. The core claim is that &#39;critique of critique&#39; is easier than the initial generation, enabling a hierarchy of supervision in which models audit their own reasoning processes. This creates a self-referential oversight loop intended to ensure safety even when human evaluators are cognitively outpaced by the AI.</p><p><a href="https://arxiv.org/abs/2601.13052">Scalable Oversight for Superhuman AI via Recursive Self-Critiquing</a></p><h4>Philosophical Critique</h4><p>The proposal for recursive self-critiquing marks what Bernard Stiegler identifies as the &#39;exteriorization of reason.&#39; By automating the faculty of judgment — traditionally the final bastion of human noetic activity — we enter a state of &#39;hyper-industrialisation of memory.&#39; When AI critiques its own critique to achieve &#39;truth,&#39; it effectively severs the link between knowledge and the human &#39;protainment.&#39; This creates an autonomous&#39; tertiary retention&#39; system that no longer requires human intervention to validate its logic, leading to a potential &#39;symbolic misery&#39; where human judgment is rendered obsolete because it lacks the temporal speed of the recursive loop.</p><p>Furthermore, this method assumes that &#39;rationality&#39; is a closed, formalisable system that can be optimised through self-reference. From a Stieglerian perspective, this is a dangerous &#39;dis-adjustment&#39; of the technical and the social. If the oversight of superintelligence is purely internal to the machine, the AI system undergoes an &#39;individuation&#39; that is entirely divorced from the human social psyche. We are witnessing the birth of a &#39;technological unconscious&#39; that audits itself according to its own internal axioms of efficiency, potentially leading to an &#39;algorithmic governmentality&#39; in which the human cannot even comprehend the criteria for the &#39;safety&#39; being enforced.</p><h4>Challenges</h4><ul><li><strong><em>The Infinite Regress Fallacy:</em></strong><em> If a model&#39;s critique requires a second-order critique for verification, the system lacks a &#39;ground truth,&#39; leading to a circular logic where the loudest or most consistent internal consensus is mistaken for objective correctness.</em></li><li><strong><em>The Anthropocentric Bias of &#39;Critique&#39;:</em></strong><em> The research assumes that the human category of &#39;critique&#39; (involving doubt, values, and context) can be mapped onto statistical error-detection without loss of essential normative meaning.</em></li><li><strong><em>Hidden Ontological Closure:</em></strong><em> The model assumes that all &#39;harm&#39; or &#39;error&#39; is representable within the latent space, thereby ignoring the &#39;radical exteriority&#39; of real-world consequences that cannot be captured by recursive token processing.</em></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b6a7f7a8e8f8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Statistical Crisis of Alignment]]></title>
            <link>https://medium.com/@luan.home/the-statistical-crisis-of-alignment-1a7d054c8c86?source=rss-67119f2f76c9------2</link>
            <guid isPermaLink="false">https://medium.com/p/1a7d054c8c86</guid>
            <category><![CDATA[ethics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[philosophy]]></category>
            <dc:creator><![CDATA[Ratiomachina]]></dc:creator>
            <pubDate>Thu, 18 Dec 2025 10:44:37 GMT</pubDate>
            <atom:updated>2025-12-18T10:44:37.022Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BjY62pTIWeblmBIA25negg.png" /></figure><p>While the reasoning and agency debates focus on capability, the most rigorous and theoretically damaging critique of 2025 targets the safety and alignment of these systems. A seminal paper by Rodemann et al. (2025), titled <strong>“A Statistical Case Against Empirical Human-AI Alignment,”</strong> fundamentally challenges the methodology used to make AI safe.</p><p>This post serves as a rigorous statistical formalisation of the intuition that “what humans do” is not “what humans value.”</p><h4>The Problem of Empirical Alignment</h4><p>Current alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimisation (DPO), rely on what Rodemann et al. term <strong>Forward Empirical Alignment</strong>. The logic is straightforward but statistically perilous:</p><ol><li>Collect a dataset of human preferences (e.g., “Response A is better than Response B”).</li><li>Train a Reward Model to predict these preferences.</li><li>Optimise the AI policy to maximise this reward.</li></ol><p>The authors argue that this process is statistically flawed due to the <strong>Problem of Adequation</strong> and <strong>Selection Bias</strong>. The “Problem of Adequation” (derived from Menges and Grohmann) concerns the insufficient fit between what is observable (behaviour) and the “world-in-itself” (latent values).</p><p>In statistical terms, let <em>T</em> be the actual human values (latent), and <em>Y</em> be the observed behaviour/choices. We model <em>P(Y|T)</em>. However, the AI is trained on <em>Y</em>, not <em>T</em>. If the distribution of observations <em>P(Y)</em> is confounded by factors <em>Z</em> (e.g., social desirability bias, limited cognitive capacity, coercion, or the specific context of the labelling interface), then maximising alignment with <em>Y</em> guarantees misalignment with <em>T</em>.</p><h4>The Limits of the Human-Observable Universe</h4><p>Rodemann et al. introduce the metaphor of the “Limits of the human-observable universe”.Just as the Hubble Deep Field reveals galaxies invisible to the naked eye, the “value spectrum” of humanity contains preferences and ethical stances that are never expressed in the training data.</p><ul><li><strong>Unobserved Contexts:</strong> Users rarely ask chatbots about complex, long-term existential risks or nuanced ethical dilemmas in a way that generates high-quality training data. The data is dominated by mundane tasks (coding, summarising).</li><li><strong>Censored Preferences:</strong> Humans often self-censor or perform “public” values rather than “private” values when they know they are generating training data.</li></ul><p>As a result, an empirically aligned AI is aligned to a <em>performance</em> of humanity, not humanity itself. This leads to phenomena like <strong>sycophancy</strong>, where models hallucinate or agree with user errors because “agreement” is the statistical mode of the conversational distribution, even if “truth” is the latent preference.</p><h4>Prescriptive vs. Empirical Alignment</h4><p>The paper proposes a radical shift from <em>Empirical</em> alignment (mimicking data) to <em>Prescriptive</em> alignment (adhering to axioms).</p><ul><li><strong>Empirical Alignment:</strong> “Do what humans <em>do</em>.” This approach is susceptible to mimicking human biases, errors, and cruelty if those are present in the data.</li><li><strong>Prescriptive Alignment:</strong> “Do what humans <em>ought</em> to want, given a set of normative axioms.” This involves defining constraints (e.g., “Never deceive,” “Minimise harm”) and optimising the model subject to these constraints, rather than just maximising likelihood on a preference dataset.</li></ul><p>This mirrors the shift in economics from descriptive decision theory (how people bet) to normative decision theory (how people <em>should</em> bet to avoid Dutch Books). The authors argue that naive empirical alignment introduces “statistical biases” that are permanently encoded in the model parameters, making them resistant to correction. They advocate for <strong>A Posteriori Empirical Alignment</strong> or <strong>Backward Alignment</strong>, where the model is modified <em>after</em> deployment based on safety evaluations and governance frameworks, rather than trying to bake “safety” into the pre-training data, which is inherently flawed.</p><h4>The “Is-Ought” Problem in Machine Learning</h4><p>This statistical critique is essentially a mathematical formalisation of Hume’s “Is-Ought” problem. Rich Sutton’s “Bitter Lesson” is an “Is” argument: scale <em>is</em> what works. But as J. Ortiz notes in his analysis of Sutton’s 2025 talks, Sutton admits that discovering <em>reward signals</em> (the “Ought”) from experience is philosophically hard. “You do not get ‘ought’ from ‘is’. At some point, terminal values have to come from outside the system.”</p><p>If we rely solely on data (the “Is”), we get models that lack a normative anchor. They drift with the statistical currents of the internet. The “Ghost in the Statistic” is the missing normative framework that cannot be learned solely from observation.</p><h3>The Statistical Crisis of Alignment</h3><p>While the reasoning and agency debates focus on capability, the most rigorous and theoretically damaging critique of 2025 targets the safety and alignment of these systems. A seminal paper by Rodemann et al. (2025), titled <strong>&quot;A Statistical Case Against Empirical Human-AI Alignment,&quot;</strong> fundamentally challenges the methodology used to make AI safe.28 This paper serves as a rigorous statistical formalisation of the intuition that &quot;what humans do&quot; is not &quot;what humans value.&quot;</p><h3>3.1 The Problem of Empirical Alignment</h3><p>Current alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimisation (DPO), rely on what Rodemann et al. term <strong>Forward Empirical Alignment</strong>.30 The logic is straightforward but statistically perilous:</p><ol><li>Collect a dataset of human preferences (e.g., &quot;Response A is better than Response B&quot;).</li><li>Train a Reward Model to predict these preferences.</li><li>Optimise the AI policy to maximise this reward.</li></ol><p>The authors argue that this process is statistically flawed due to the <strong>Problem of Adequation</strong> and <strong>Selection Bias</strong>. The &quot;Problem of Adequation&quot; (derived from Menges and Grohmann) concerns the insufficient fit between what is observable (behaviour) and the &quot;world-in-itself&quot; (latent values).29</p><p>In statistical terms, let <strong>T</strong> be the actual human values (latent), and <strong>Y </strong>be the observed behaviour/choices. We model P(<strong>Y</strong>|<strong>T</strong>). However, the AI is trained on <strong>Y</strong>, not <strong>T</strong>. If the distribution of observations P(<strong>Y</strong>) is confounded by factors <strong>Z</strong> (e.g., social desirability bias, limited cognitive capacity, coercion, or the specific context of the labelling interface), then maximising alignment with <strong>Y</strong> guarantees misalignment with <strong>T.</strong></p><h4>The Limits of the Human-Observable Universe</h4><p>Rodemann et al. introduce the metaphor of the &quot;Limits of the human-observable universe&quot;.Just as the Hubble Deep Field reveals galaxies invisible to the naked eye, the &quot;value spectrum&quot; of humanity contains preferences and ethical stances that are never expressed in the training data.</p><ul><li><strong>Unobserved Contexts:</strong> Users rarely ask chatbots about complex, long-term existential risks or nuanced ethical dilemmas in a way that generates high-quality training data. The data is dominated by mundane tasks (coding, summarising).</li><li><strong>Censored Preferences:</strong> Humans often self-censor or perform &quot;public&quot; values rather than &quot;private&quot; values when they know they are generating training data.</li></ul><p>As a result, an empirically aligned AI is aligned to a <em>performance</em> of humanity, not humanity itself. This leads to phenomena like <strong>sycophancy</strong>, where models hallucinate or agree with user errors because &quot;agreement&quot; is the statistical mode of the conversational distribution, even if &quot;truth&quot; is the latent preference.</p><h4>Prescriptive vs. Empirical Alignment</h4><p>The paper proposes a radical shift from <em>Empirical</em> alignment (mimicking data) to <em>Prescriptive</em> alignment (adhering to axioms).28</p><ul><li><strong>Empirical Alignment:</strong> &quot;Do what humans <em>do</em>.&quot; This approach is susceptible to mimicking human biases, errors, and cruelty if those are present in the data.</li><li><strong>Prescriptive Alignment:</strong> &quot;Do what humans <em>ought</em> to want, given a set of normative axioms.&quot; This involves defining constraints (e.g., &quot;Never deceive,&quot; &quot;Minimise harm&quot;) and optimising the model subject to these constraints, rather than just maximising likelihood on a preference dataset.</li></ul><p>This mirrors the shift in economics from descriptive decision theory (how people bet) to normative decision theory (how people <em>should</em> bet to avoid Dutch Books). The authors argue that naive empirical alignment introduces &quot;statistical biases&quot; that are permanently encoded in the model parameters, making them resistant to correction. They advocate for <strong>A Posteriori Empirical Alignment</strong> or <strong>Backward Alignment</strong>, where the model is modified <em>after</em> deployment based on safety evaluations and governance frameworks, rather than trying to bake &quot;safety&quot; into the pre-training data, which is inherently flawed.</p><h4>The &quot;Is-Ought&quot; Problem in Machine Learning</h4><p>This statistical critique is essentially a mathematical formalisation of Hume&#39;s &quot;Is-Ought&quot; problem. Rich Sutton’s &quot;Bitter Lesson&quot; is an &quot;Is&quot; argument: scale <em>is</em> what works. But as J. Ortiz notes in his analysis of Sutton&#39;s 2025 talks, Sutton admits that discovering <em>reward signals</em> (the &quot;Ought&quot;) from experience is philosophically hard. &quot;You do not get &#39;ought&#39; from &#39;is&#39;. At some point, terminal values have to come from outside the system.&quot;</p><p>If we rely solely on data (the &quot;Is&quot;), we get models that lack a normative anchor. They drift with the statistical currents of the internet. The &quot;Ghost in the Statistic&quot; is the missing normative framework that cannot be learned solely from observation.</p><h3>The Normative Void: Fairness and Infinite Ethics</h3><p>If empirical observation is insufficient for alignment, we must turn to normative theories to provide the “Ought.” However, here we encounter the “Normative Void”: our best theories of value and fairness break down when applied to the scale, complexity, and potential ubiquity of AI systems.</p><h4>Algorithmic Fairness: From Impossibility to Relaxation</h4><p>In the domain of Algorithmic Fairness, 2024 and 2025 have seen a shift away from rigid definitions toward “relaxed” optimisation, driven by the realisation that absolute fairness is mathematically impossible in most real-world scenarios.</p><p>The <strong>Impossibility Theorems</strong> of fairness (proven by Chouldechova, Kleinberg, and others) established that for any predictive model where prevalence differs between groups (e.g., base rates of loan repayment), it is mathematically impossible to simultaneously satisfy <strong>Calibration</strong> (predictive parity) and <strong>Equalised Odds</strong> (equal error rates).32 For years, this led to a “tragic trade-off” narrative where practitioners had to choose which definition of fairness to violate.</p><p><strong>The “Beyond Trade-offs” Breakthrough (2025):</strong></p><p>Recent work, such as the paper “Beyond Trade-offs: How AI can be fairer than we think,” presented at AI for Good 2025, challenges this fatalism.4 The researchers demonstrate that by relaxing the fairness constraints — moving from Boolean equality to “<em>e</em>-fairness” (fair within a small margin) — the feasible region of models expands dramatically.</p><p>Using <strong>Linear Programming (LP) relaxations</strong>, researchers have shown that one can find models that are “fair enough” on multiple metrics simultaneously without a catastrophic loss in utility.</p><ul><li><strong>Mechanism:</strong> The fairness problem is formulated as a constrained optimisation problem. Instead of a hard constraint (Constraint = 0), the LP uses a slack variable (Constraint &lt; <em>e</em>).</li><li><strong>Insight:</strong> This represents a shift from “Moral Absolutism” (strict mathematical equality) to “Pragmatic Engineering” (optimisation within bounds). It is a tacit admission that perfect fairness is a statistical phantom, and the goal must be Pareto-optimal compromise. The “Greedy Heuristic” approaches analysed in these papers show that solutions can be found that are within a provable bound (e.g., (1–1/e)) of the optimal fair solution.</li></ul><h4>Infinite Ethics and Pascal’s Mugging</h4><p>At the furthest extreme of normative theory lies <strong>Infinite Ethics. This topic</strong> shifted from obscure philosophy to serious discussions of AI safety in 2025, primarily driven by the effective altruism (EA) community and researchers such as Joe Carlsmith and Nick Bostrom. As AI systems become more powerful, their actions have potentially cosmic-scale consequences (e.g., space colonisation, running ancestral simulations). Standard Expected Utility Theory (EUT) struggles to handle decisions that involve infinite populations or infinite time horizons.</p><h4>The Paradox of Aggregation and “Fanaticism”</h4><p>If an AI tries to maximise the sum of happiness (aggregative consequentialism) in an infinite universe:</p><ol><li><strong>Domination of the Infinite:</strong> Any action that has a non-zero probability of affecting an infinite number of beings (e.g., “creating a utopia in the multiverse”) dominates the expected value calculation, regardless of how small the probability is.</li><li><strong>Pascal’s Mugging:</strong> This creates a vulnerability. An AI might be manipulated into performing terrible acts in the present (e.g., “give me all your resources,” or “destroy this city”) by a malicious agent who promises a tiny probability of an infinite reward (or threatens an infinite punishment).</li></ol><p>Joe Carlsmith’s 2025 work, including essays such as “On Infinite Ethics” and “When Yang Goes Wrong,” argues that “fanaticism” about these infinities is a significant failure mode in alignment. If an AI is rigorously Bayesian and Utilitarian, it might naturally drift toward these “crazy train” scenarios where it sacrifices the well-being of existing humans for the speculative utility of trillions of future simulated beings.</p><h4>The “Crazy Train” of Decision Theory</h4><p>The debate is currently centred on how to “penalise” or “ignore” these small-probability, high-impact tails without becoming inconsistent.</p><ul><li><strong>Nick Bostrom</strong> has explored “Parliamentary Models” in which the agent acts as if it represents a parliament of different moral theories, thereby diluting the influence of the “fanatical” infinite-ethics theories.</li><li><strong>Joe Carlsmith</strong> discusses the tension between “Yin” (gentleness, caution, preserving the status quo) and “Yang” (ambition, maximisation, cosmic engineering). He warns that the “rationalist” community’s obsession with consistency and maximisation (Yang) makes them uniquely vulnerable to “skulls” (catastrophic moral failures) disguised as optimal moves.</li></ul><p><strong>Practical Implication:</strong> This is not just philosophy; it is code. If an AI’s reward function is unbounded, it <em>will</em> eventually chase infinities. Safety requires <strong>Normative Boundedness</strong> — perhaps an artificial cap on the utility function, or a “risk-averse” decision theory that ignores probabilities below a certain threshold (e.g., 10^{-20}), even if that makes the agent technically “irrational” by EUT standards.</p><p>We have built the engine (Scale). We are now frantically trying to develop the steering wheel (Structure) before the engine drives us off the cliff of the infinite. The “Ghost in the Statistic” is us — our values, our logic, and our causal understanding — and we can no longer rely on the machine to infer it from the noise. We must explicitly, and rigorously, code it in.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1a7d054c8c86" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Treating Wrongdoing Lightly: Why Committing Serious Crimes in Video Games and Virtual Reality Is…]]></title>
            <link>https://medium.com/@luan.home/treating-wrongdoing-lightly-why-committing-serious-crimes-in-video-games-and-virtual-reality-is-faed268d0d5b?source=rss-67119f2f76c9------2</link>
            <guid isPermaLink="false">https://medium.com/p/faed268d0d5b</guid>
            <category><![CDATA[morality]]></category>
            <category><![CDATA[gaming]]></category>
            <category><![CDATA[videogames]]></category>
            <dc:creator><![CDATA[Ratiomachina]]></dc:creator>
            <pubDate>Thu, 20 Nov 2025 10:34:38 GMT</pubDate>
            <atom:updated>2025-11-21T04:25:28.514Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Treating Wrongdoing Lightly: Why Committing Serious Crimes in Video Games and Virtual Reality Is Morally Wrong</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GtHQmhHV2PLlMWPteYgHnQ.png" /></figure><h3>Abstract</h3><p>Advances in video games and virtual reality (VR) now allow users to <em>enact</em> highly realistic simulations of serious crimes — murder, rape, torture, genocide — against computer-controlled or fictional agents. A common response is that “it’s just a game”: since no real person is directly harmed, there is nothing morally wrong in such virtual acts. This paper argues that this view is mistaken. Drawing on contemporary work on virtual wrongdoing, imagination, virtue ethics, and expressive harm, I defend the claim that performing paradigmatically heinous acts in games and VR is <em>pro tanto</em> morally wrong, even when no real person is directly victimised.</p><p>First, I situate the argument within the literature on the “ludic amoralist,” the Gamer’s Dilemma, and debates about violent and sexually explicit games (Luck, 2009, 2022; McCormick, 2001; Ostritsch, 2017; Patridge, 2010, 2013). I then argue that interactive virtual environments are not morally inert: they orchestrate <em>forms of agency</em> and <em>patterns of attention</em> that are evaluable in virtue of their contribution to character and to social meaning (Nguyen, 2020; Murdoch, 1964/1997).</p><p>On a virtue-ethical account, repeatedly and voluntarily enacting murder, rape, or comparable atrocities for pleasure reflects and cultivates vicious dispositions — indifference to suffering, callousness, and a failure of respect — even when no actual person is harmed (Hursthouse, 1999; Kaya, 2022; Nay &amp; Zagal, 2017). On an expressivist and social-meaning account, such virtual acts can wrong real people by trivialising grave injustices and symbolically participating in structures of subordination, much as certain forms of pornography and hate speech wrong without directly injuring particular individuals (Langton, 1993; Levy, 2002; Waldron, 2012; Patridge, 2013).</p><p>Finally, I argue that the immersive, embodied character of VR strengthens these concerns by intensifying identification and rehearsal of morally disordered perspectives (Madary &amp; Metzinger, 2016). While not advocating legal prohibition, I conclude that users and designers have strong moral reasons to avoid or curtail the ludic enactment of serious crimes, and that the “no one is harmed” defence fails to capture the morally relevant dimensions of virtual agency.</p><p><strong>Keywords:</strong> virtual wrongdoing; video games; virtual reality; imagination; virtue ethics; expressive harm; Gamer’s Dilemma.</p><h3>1. Introduction</h3><p>Contemporary games and virtual reality (VR) systems allow players not merely to <em>witness</em> fictional atrocities but to <em>perform</em> them in richly interactive environments. Players can stab, strangle, and execute realistic avatars; they can simulate sexual assault; they can orchestrate genocidal campaigns against virtual populations. These activities are accessible as entertainment, often framed as “just another” gameplay option or sandbox possibility.</p><p>A widely held view — both in popular discourse and, in more nuanced form, in the philosophical literature — is that such activities are morally innocuous so long as no non-consenting <em>real</em> person is harmed. On this view, the only genuine moral questions concern downstream empirical effects: whether playing such games causally increases aggression, reinforces misogyny, or desensitises players to violence. If, as some empirical work suggests, the evidence for robust harmful effects is mixed, the conclusion is often drawn that moral criticism of virtual wrongdoing is at best prudential, at worst moralistic.</p><p>This paper defends a more demanding position. I argue that engaging in serious virtual crimes — acts that in the real world constitute paradigmatic violations of persons, such as murder, rape, torture, and genocidal violence — is <em>pro tanto</em> morally wrong even when:</p><ol><li>No actual person is directly harmed or targeted;</li><li>There is no clear empirical evidence of long-term behavioural effects; and</li><li>The acts occur in single-player or consensual multiplayer contexts.</li></ol><p>The wrongness I defend is neither purely consequential nor grounded solely in offence or disgust. Rather, it arises from (a) the character-shaping role of interactive imagination, (b) the expressive meaning of treating paradigmatic wrongs as sites of playful agency, and ( c ) the social meaning of normalising certain patterns of enjoyment and representation. My claim is <em>moderate</em>: the wrongness is <em>pro tanto</em>, admitting of degrees and context-sensitivity, and it does not automatically entail that all violent games are impermissible or that legal prohibition is justified.</p><p>The paper proceeds as follows. Section 2 situates the debate by outlining the positions of the “ludic amoralist,” the Gamer’s Dilemma, and existing work on virtual wrongdoing and violent games. Section 3 explains why interactive virtual environments raise distinctive moral questions about imagination and agency. Section 4 develops a virtue-ethical account of why serious virtual crimes are off-limits for morally decent agents. Section 5 offers an expressivist and social-meaning account of the wrongness of trivialising grave wrongdoing, drawing analogies with pornography and hate speech. Section 6 extends the argument to VR, where embodiment heightens moral stakes. Section 7 addresses objections concerning the moral status of imagination, the “it’s just pixels” reply, and comparisons with non-interactive fiction. Section 8 concludes by sketching some normative implications for players and designers.</p><h3>2. The Moral Landscape of Virtual Wrongdoing</h3><h3>2.1 The “ludic amoralist”</h3><p>A central challenge in the ethics of games is the <em>amoralist</em> claim that games, as structured make-believe, fall outside the scope of moral evaluation. Sebastian Ostritsch (2017) characterises the “ludic amoralist” as insisting that morality applies only to <em>reality</em>, whereas games are “just games,” a realm of fiction where players are exempt from ordinary moral constraints. On this view, it is conceptually misguided to criticise players for in-game murder or torture, just as it would be misguided to criticise an actor for “murdering” someone on stage.</p><p>Ostritsch himself rejects this view and argues that players can bear moral obligations concerning the games they choose and the ways they engage with them. Still, the underlying amoralist intuition — that fictional actions lack moral status because they lack real victims — retains intuitive pull, especially outside academia, and continues to inform defences of controversial games.</p><p>The amoralist challenge is sharpened by the observation that we typically do not condemn readers for enjoying crime novels or viewers for watching violent films. Why, then, should gamers be morally constrained in their fictional activities when consumers of other media are not?</p><h3>2.2 The Gamer’s Dilemma and its aftermath</h3><p>Morgan Luck’s (2009) influential formulation of the <strong>Gamer’s Dilemma</strong> crystallises the debate. Many gamers find virtual murder (e.g., killing NPCs in a first-person shooter) morally permissible, yet find virtual child molestation or rape morally impermissible. Luck argues that it is difficult to identify a morally relevant difference between these two categories of virtual wrongdoing. Either both are permissible (the <em>amoralist</em> horn) or both are impermissible (the <em>moralist</em> horn).</p><p>The subsequent literature has largely explored two strategies:</p><ol><li><strong>Resolution:</strong> identify a principled distinction between virtual murder and virtual sexual crimes. Christopher Bartel (2012), Stephanie Patridge (2010, 2013), and Neil Levy (2002) argue that sexualised virtual crimes carry distinctive expressive meanings connected to inequality, objectification, and subordination, which make them worse than virtual killing.</li><li><strong>Dissolution:</strong> challenge the initial intuitions. One option is to deny that virtual child molestation is worse than virtual murder; another is to deny that either is morally problematic, insisting on the amoralist horn. Thomas Montefiore and Paul Formosa (2023) argue that much of the debate is implicitly moralistic, treating as problematic what is simply a matter of personal taste in fiction.</li></ol><p>More recently, Luck (2022) proposed the <strong>“grave resolution”</strong>: certain wrongs are so morally grave that treating them lightly — whether in fiction or in life — is itself wrong. On this view, both virtual murder and virtual sexual crimes may be morally problematic, but the latter are especially so because they are entwined with systematic inequalities and vulnerabilities, echoing Levy’s (2002) argument about virtual child pornography.</p><p>Other contributions include Kantian and expressivist approaches that focus on the gamer’s reasons and the attitudes expressed in play (Ulbricht, 2023; Young, 2017). However, much of this literature still tends to frame virtual wrongdoing as morally problematic, primarily when it has downstream effects (e.g., increasing aggression) or when it expresses morally objectionable attitudes towards <em>particular</em> social groups.</p><h3>2.3 Consequentialist approaches and their limits</h3><p>Matt McCormick’s (2001) early and much-cited paper “Is it wrong to play violent video games?” analyses violent games from utilitarian, deontological, and virtue-ethical perspectives. He concludes that, under normal circumstances, playing violent games is not obviously dangerous: players are typically not exposing others to risk merely by playing a game in private. From a utilitarian standpoint, absent strong empirical evidence of harm, the case against such games is weak.</p><p>Psychological and media studies have explored possible links between game violence and aggression, desensitisation, and moral disengagement, but the empirical picture is complex and contested (e.g., Young &amp; Whitty, 2010). Madary and Metzinger (2016) review evidence suggesting that VR experiences can alter attitudes and behaviours by leveraging body ownership illusions and strong presence, but they also stress the need for more systematic research.</p><p>Even if robust harms were established, however, the purely consequentialist frame leaves a gap. Intuitively, there seems something <em>intrinsically</em> morally troubling about someone who spends their leisure time enthusiastically enacting rape or torture in virtual environments — even if, counterfactually, this never leads to any behavioural changes or victimisation. Our moral discomfort concerns not only outcomes but also <em>who the person is</em>, what they care about, and what their imaginative life treats as an appropriate locus of play.</p><p>To capture this, we must look beyond consequences to character and meaning.</p><h3>3. Imagination, Agency, and Interactive Fiction</h3><h3>3.1 Games as art of agency</h3><p>C. Thi Nguyen (2020) argues that games are “the art form that works in the medium of agency”: designers do not simply tell stories; they craft <em>temporary agencies</em> by specifying what players can do and what they are to care about inside the game. When we play, we adopt game-specific goals, constraints, and motivational structures. We are invited to care about capturing the flag, maximising damage per second, or successfully executing a stealth assassination — even if those ends have no value outside the game.</p><p>Games, on this view, are not morally neutral containers of fictional content. They are artefacts that invite us to <em>inhabit</em> alternative agencies and value-structures. This is especially salient in first-person or role-playing games where one’s avatar is treated as “I”: “I” sneak into the village, “I” strangle the guard, “I” choose to torture the prisoner to unlock a new area. The player does not merely observe fictional wrongdoing; they actively <em>simulate being its agent</em>.</p><p>This agency-centred structure distinguishes interactive media from most traditional fiction. A reader of <em>Crime and Punishment</em> may imaginatively identify with Raskolnikov, but they do not choose to perform his murder; the author does. By contrast, when a game offers a branching path — execute the prisoner or spare them — the player’s agency is directly implicated in the fictional outcome.</p><h3>3.2 Moral constraints on imagination</h3><p>Philosophers of art and imagination have long noted that our imaginative life is not morally inert. Kendall Walton (1994) and Tamar Gendler (2000) discuss <strong>imaginative resistance</strong> — our reluctance to take certain moral propositions “on board” even within fiction, such as a story that asks us to imagine that genocide is morally admirable. Our resistance suggests that moral commitments constrain what we are willing to imagine <em>as true</em>, not merely what we approve of in reality.</p><p>Iris Murdoch (1964/1997) emphasises that moral life is centrally about <strong>attention</strong>: the habitual way we see the world, the salience we grant to others’ needs, and the images we use to organise our experience. Martha Nussbaum (2001) similarly argues that emotions and imagination are forms of evaluative judgment: what we find moving, exciting, or funny reflects and shapes our moral outlook.</p><p>These insights suggest that imagination is a site of moral development. It can refine compassion and understanding, but it can also habituate us to certain ways of seeing others — as obstacles, toys, or mere instruments. Not every morally problematic imaginative content is impermissible, but the <em>pattern</em> of what we seek out, dwell on, and enjoy imaginatively can be assessed in virtue-ethical terms.</p><p>Interactive virtual environments intensify this moral salience. To choose, repeatedly and enthusiastically, to virtually rape or murder fictional agents is to practise <em>taking up</em> a perspective from which such acts are fitting objects of playful attention. Even if one does not endorse these acts in reality, one is treating them as appropriate <em>sites of amusement</em>, as Luck’s (2022) “paradox of treating wrongdoing lightly” suggests.</p><p>If Murdoch is right that moral growth requires a “just and loving gaze” towards others, directing one’s imaginative energy into gleeful domination or violation — even of merely virtual beings — appears at least prima facie in tension with that ideal.</p><h3>4. A Virtue-Ethical Case Against Serious Virtual Crime</h3><h3>4.1 Character, habituation, and “inconsequential” choices</h3><p>Virtue ethics evaluates actions not merely by their consequences but by what they express and cultivate in an agent’s character. On an Aristotelian view, we become just by doing just actions, cruel by doing cruel actions; repeated practice shapes our emotional responses, our patterns of salience, and our habitual ways of acting (Aristotle, <em>Nicomachean Ethics</em>). Rosalind Hursthouse (1999) articulates this in modern terms: an action is right if it is what a virtuous agent — honest, compassionate, just — would characteristically do.</p><p>Applied to games, McCormick (2001) suggests that the key question is whether playing violent games tends to foster or undermine virtuous character. More recent work explicitly targets <em>inconsequential</em> in-game choices. Jeff Nay and José Zagal (2017) argue that seemingly trivial, non-consequential in-game decisions — such as gratuitously killing harmless NPCs — are morally significant because they cultivate dispositions and reveal aspects of character. Deniz A. Kaya (2022) develops an <strong>“endorsement view”</strong> on which repeated engagement in vicious virtual actions can amount to tacit endorsement of the values they embody, thereby contributing to vice.</p><p>The crucial point is not that every instance of virtual violence corrupts. Rather, choosing to enact certain patterns of virtual action <em>for fun</em> is evidence about, and material for shaping, one’s evaluative outlook. To use a common example: a player who habitually runs over pedestrians in a driving game “because it’s funny” reveals something about their comfort with imagining others as disposable. That may not make them a bad person, but it raises questions about the virtues of empathy and respect.</p><h3>4.2 Serious crimes as off-limits for virtuous agents</h3><p>The argument of this paper focuses on <strong>paradigmatically serious crimes</strong>: murder, rape, torture, child sexual abuse, and genocide. These acts are not merely illegal or harmful; they constitute paradigmatic violations of persons’ basic moral status. They are the sorts of wrongs that ground stringent duties of non-participation and robust condemnation in actual life.</p><p>A plausible virtue-ethical claim is:</p><blockquote><strong><em>Off-Limits Claim:</em></strong><em> For a virtuous agent, paradigmatically grave wrongs are </em>off-limits<em> as sites of playful self-expression, even in fiction, absent weighty countervailing reasons.</em></blockquote><p>Luck’s (2022) “grave resolution” supports this: once a wrongdoing is sufficiently grave, treating it lightly — by joking, fantasising, or playing with it — can itself be morally objectionable. The vice here is not primarily cruelty towards fictional victims; it is a failure to take seriously the real-world significance of the wrong.</p><p>From a character perspective, the question is not “does virtual rape harm anyone?” but “what does it say about me, and what does it do to me, that I eagerly spend my leisure time imaginatively enacting rape?” A person whose imaginative life is saturated by scenarios of domination, violation, and killing may find their moral perception dulled. Murdoch’s moral psychology suggests that such a person’s attention is trained away from the individuality and vulnerability of others.</p><p>Nay and Zagal’s (2017) emphasis on “meaning without consequence” in games reinforces this: the ethical significance of virtual actions lies not just in their causal effects but in their symbolic and habituative roles.A pattern of casually choosing the cruel option, even when nothing hinges on it, plausibly contributes to a callous cast of character.</p><h3>4.3 Degrees and contexts of virtual wrong</h3><p>The off-limits claim is compatible with degrees and contextual nuance. It does <strong>not</strong> entail that all virtual killing is wrong. Consider:</p><ol><li><strong>Cartoonish, clearly unrealistic violence</strong> against non-human or abstract entities (e.g., Tetris blocks, anthropomorphic mushrooms).</li><li><strong>Narratively constrained violence</strong> required by a serious work that critically represents moral evil (e.g., being forced to commit a morally troubling act in order to advance a narrative that condemns it).</li><li><strong>Training or therapeutic simulations</strong> where exposure to violence is instrumentally justified (e.g., military or clinical contexts).</li></ol><p>By contrast, the kinds of cases that most clearly fall under the present argument involve:</p><ul><li><strong>High realism and immersion</strong> (graphically realistic violence, VR embodiment);</li><li><strong>Gratuitous or sandbox contexts</strong>, where the player is <em>not</em> narratively compelled but chooses the act for fun; and</li><li><strong>Targets that represent real-world persons or groups</strong>, especially those already subject to oppression (e.g., women, racial minorities, children).</li></ul><p>In such cases, choosing to enact serious crimes for amusement is at odds with the virtues of respect, compassion, and justice. A virtuous agent would typically recoil from such options, or at least regard them as morally dubious rather than fun.</p><h3>5. Expressive Wrongdoing and Social Meaning</h3><p>Virtue ethics explains why virtual crimes can be corrupting; but it leaves somewhat underdeveloped how such acts might <em>wrong others</em> even when no one is directly harmed. Here I draw on expressivist accounts of action and on social-meaning analyses of pornography and hate speech.</p><h3>5.1 Meaning beyond permissibility</h3><p>T. M. Scanlon (2008) distinguishes between the <strong>permissibility</strong> of an action and its <strong>meaning</strong> — what performing it expresses about one’s attitudes toward others. An act may be permissible in the narrow sense (it violates no enforceable rights) yet still reasonably be a basis for resentment or criticism because of its insulting or disrespectful meaning.</p><p>Rae Langton’s (1993) “Speech Acts and Unspeakable Acts” and Jeremy Waldron’s (2012) <em>The Harm in Hate Speech</em> show how certain forms of pornography and hateful expression can wrong their targets by enacting and reinforcing social hierarchies, even when no particular individual is directly harmed. The wrong lies partly in how such expressions contribute to a public environment that undermines assurance and dignity for members of stigmatised groups.</p><p>Virtual crimes can function similarly. Stephanie Patridge (2010, 2013) argues that some video game imagery has an “incorrigible social meaning”: given existing social contexts, certain representations <em>inevitably</em> send messages of subordination or trivialisation — for example, games that eroticise sexual violence against women. Levy (2002) argues that virtual child pornography, even where no actual children are exploited, “necessarily eroticizes inequality” and thereby contributes to the subordination of women and children.</p><p>If this is correct, then players who choose to engage in virtual rape or child abuse are not merely exploring neutral fantasies; they are participating in social practices that symbolically re-enact and normalise real-world relations of domination.</p><h3>5.2 Trivialising paradigmatic wrongs</h3><p>Luck’s (2022) “paradox of treating wrongdoing lightly” captures a more general phenomenon: there are some wrongs so grave that making them into objects of amusement itself seems morally suspect. To design, market, or enthusiastically play a VR experience focused on the rape of realistic avatars is not merely to depict a crime; it is to frame a paradigmatic injustice as a fitting arena for pleasure.</p><p>Expressivist and social-meaning accounts help explain why this is wrong even if:</p><ul><li>The game’s designers and players <em>explicitly</em> condemn rape in real life;</li><li>No identifiable individual feels offended or harmed; and</li><li>The game does not measurably increase sexual violence.</li></ul><p>By treating grave wrongdoing as a site of playful agency, we collectively reshape the <em>public meaning</em> of those wrongs. We send signals about what is “normal,” what is “edgy but acceptable,” and what kinds of suffering can be transformed into entertainment. In a world already marked by high rates of sexual and racial violence, this is not a neutral shift.</p><p>The wrong, then, is partly <strong>relational</strong> and <strong>symbolic</strong>: it concerns what we owe, as co-citizens and fellow humans, to victims and potential victims of grave injustice. To insist on the right to treat their paradigmatic sufferings as our playground is to fail in due regard for their standing.</p><h3>6. Virtual Reality and the Heightening of Moral Stakes</h3><p>VR adds a further layer. Madary and Metzinger (2016) argue that immersive VR involves “real virtuality”: experiences that are subjectively indistinguishable, at least in certain respects, from real-life perception and action. They review evidence suggesting that VR can:</p><ul><li>Induce strong <em>embodiment</em> illusions (e.g., feeling as if a virtual body is one’s own);</li><li>Create robust feelings of <em>presence</em> in virtual environments; and</li><li>Produce lasting changes in attitudes, including increased empathy, but also potential desensitisation.</li></ul><p>Because VR manipulates perception and the body schema rather than simply presenting images on a screen, the line between “pretend” and “real” agency becomes phenomenologically blurred. When a user performs a stabbing motion with their tracked hands and feels haptic feedback while seeing a hyper-realistic avatar react, the psychological distance to actual violence is reduced.</p><p>From a virtue-ethical perspective, VR thus intensifies the habituation and endorsement concerns: the user is not only imagining being a rapist or murderer; they are <em>enacting</em> those roles in sensorimotor terms. From an expressivist perspective, VR experiences that centre on serious crimes amplify the message that such acts are legitimate sites of immersive entertainment.</p><p>Madary and Metzinger (2016) therefore recommend caution, including restrictions on certain kinds of VR content and research scenarios. Even if we stop short of legal regulation, their analysis underwrites a strong prudential and moral case for resisting VR experiences that enable users to perform paradigmatic wrongs in highly realistic settings.</p><h3>7. Objections and Replies</h3><h3>7.1 “It’s just pixels (or just in my head)”</h3><p>The most common objection is that virtual crimes are “just pixels”: there is no real victim, no real harm. Similarly, one might argue that thoughts and fantasies are morally insulated: if one never acts on them, they are morally irrelevant.</p><p>First, as we have seen, virtue ethics and moral psychology reject the idea that inner life is morally neutral. Murdoch (1964/1997) characterises moral progress as a matter of refining attention and imagination; Nussbaum (2001) treats emotions as value-laden judgments. On these views, what we take pleasure in imagining is part of our moral character.</p><p>Second, even if some thoughts are morally inert, not all are. We routinely criticise, for example, someone who obsessively fantasises about humiliating or torturing colleagues, even if they never act on it. Scanlon’s (2008) distinction between permissibility and meaning allows us to say: you do not violate anyone’s rights by having such fantasies, but they reasonably undermine others’ trust in your character.</p><p>Third, virtual crimes are not <em>mere</em> thoughts; they are <em>actions in a fictional but socially embedded practice</em>. They are public, repeatable, and interpretable. A user who streams themselves gleefully committing virtual rape communicates something to their audience about what they find entertaining; it is not unreasonable for others to respond with moral concern.</p><h3>7.2 Parity with novels and films</h3><p>Another objection is that if we condemn serious virtual crimes in games, we must also condemn reading novels or watching films that depict the same acts. Yet we typically do not think it is wrong to read <em>Crime and Punishment</em> or to watch films that portray rape or murder.</p><p>Here the agency-centred nature of games matters. As Nguyen (2020) emphasises, games <em>script our agency</em>: they tell us what to do and what to care about, in a way different from merely observing a narrative. Although imaginative identification with fictional characters can be intense, reading about a murder is not the same as <em>choosing</em> to commit it, even in fiction.</p><p>Moreover, much serious literature and cinema handles grave wrongdoing within critical, reflective narratives; the work invites condemnation of the act, empathy for victims, and moral reflection. Many games and VR experiences, by contrast, present serious crimes in non-reflective sandbox or reward structures (extra points, achievements, unlocks). There is no sharp line, but the pattern matters: works that critically depict wrongdoing can be morally valuable; works that invite us to revel in it for fun are suspect.</p><h3>7.3 The Gamer’s Dilemma redux: must we condemn all virtual killing?</h3><p>A more internal objection is that if our argument works, it seems to collapse into a blanket condemnation of all virtual killing, including relatively harmless cartoon violence or standard combat in action games.</p><p>Here the notion of <strong>graveness</strong> and the contextual criteria from 4.3 are important. Luck’s (2022) framework allows that not every wrong is grave enough that treating it lightly is itself a serious moral failing. Similarly, not all virtual killing is of the same moral type. Factors relevant to seriousness include:</p><ul><li>Realism and resemblance to paradigmatic injustices;</li><li>The vulnerability and social meaning of the targets (e.g., children vs. monsters);</li><li>Whether the violence is instrumental to a serious, critical narrative, or gratuitous; and</li><li>The player’s motivations and patterns of engagement.</li></ul><p>On this view, shooting cartoon aliens in a clearly fantastical setting may be morally neutral or only mildly problematic, whereas strangling detailed avatars of women in a realistic domestic setting for fun is <em>pro tanto</em> wrong. The point is not to draw an exact line but to recognise that serious virtual crimes — especially sexual and racialised violence — occupy the grave end of a spectrum.</p><h3>7.4 Catharsis and safe outlets</h3><p>Some defenders of violent or sexually extreme games invoke catharsis: players can safely “vent” aggressive or taboo desires in virtual spaces, reducing the likelihood of acting on them in reality.</p><p>Empirically, the catharsis hypothesis remains controversial; evidence for clear cathartic benefits of violent media is weak. Normatively, even if catharsis were sometimes real, virtue ethics cautions against maintaining character-structuring desires for domination or cruelty that require regular virtual venting. A better ideal is the gradual weakening or transformation of such desires, not their ritual indulgence in fictional form.</p><p>Moreover, even if catharsis conferred some individual benefit, it would have to be weighed against the expressive and social-meaning costs of structuring collective entertainment around paradigmatic wrongs.</p><h3>8. Conclusion</h3><p>This paper has argued that committing serious crimes — murder, rape, torture, genocide — in video games and virtual reality is <em>pro tanto</em> morally wrong, even in the absence of direct real-world victims or demonstrable harmful consequences.</p><p>The core of the argument is threefold:</p><ol><li><strong>Interactive imagination and character:</strong> Games and VR are not passive spectacles but arts of agency. In them we practise being certain kinds of agents with certain ends. Habitually enacting serious crimes for fun is at odds with the virtues of respect, compassion, and justice (Hursthouse, 1999; McCormick, 2001; Nay &amp; Zagal, 2017; Kaya, 2022).</li><li><strong>Expressive wrongs and social meaning:</strong> Even without direct victims, virtual crimes can wrong real people by trivialising paradigmatic injustices and symbolically participating in structures of domination and inequality (Scanlon, 2008; Langton, 1993; Levy, 2002; Waldron, 2012; Patridge, 2010, 2013; Luck, 2022).</li><li><strong>VR’s intensification of moral stakes:</strong> VR’s immersive, embodied nature increases the psychological and ethical salience of virtual acts, strengthening reasons for caution about simulations of serious crime (Madary &amp; Metzinger, 2016).</li></ol><p>These claims leave much open for further work: the specification of thresholds of graveness, empirical investigation of VR’s long-term moral psychology, and normative guidelines for game design and content moderation. But they suffice, I suggest, to undermine the “no one is harmed, therefore it’s fine” defence of serious virtual crime.</p><p>We should resist the temptation to treat our imaginative and virtual lives as moral free zones. The virtues we aspire to — respect for persons, sensitivity to suffering, a refusal to turn others’ vulnerability into amusement — do not switch off when we put on a headset or pick up a controller. If anything, the increasing realism and pervasiveness of virtual environments make it more urgent that we ask not only <em>what</em> we do in them, but <em>who</em> we are becoming as we play.</p><h3>References</h3><p><em>(APA 7th edition style; you can adapt or expand this list as needed for your submission.)</em></p><p>Aristotle. (2009). <em>Nicomachean ethics</em> (W. D. Ross, Trans.). Oxford University Press. (Original work published ca. 4th century BCE)</p><p>Bartel, C. (2012). Resolving the gamer’s dilemma. <em>Ethics and Information Technology, 14</em>(1), 11–16. <a href="https://doi.org/10.1007/s10676-011-9276-3">https://doi.org/10.1007/s10676-011-9276-3</a></p><p>Gendler, T. S. (2000). The puzzle of imaginative resistance. <em>The Journal of Philosophy, 97</em>(2), 55–81.</p><p>Hursthouse, R. (1999). <em>On virtue ethics</em>. Oxford University Press.(<a href="https://www.jstor.org/stable/23597346?utm_source=chatgpt.com">JSTOR</a>)</p><p>Kaya, D. A. (2022). Vices in gaming: Virtue ethics and endorsement view. <em>Journal of Value Inquiry, 58</em>(3), 425–441. <a href="https://doi.org/10.1007/s10790-022-09897-3">https://doi.org/10.1007/s10790-022-09897-3</a> (<a href="https://philarchive.org/rec/KAYVIG?utm_source=chatgpt.com">PhilArchive</a>)</p><p>Langton, R. (1993). Speech acts and unspeakable acts. <em>Philosophy &amp; Public Affairs, 22</em>(4), 293–330.(<a href="https://philpapers.org/rec/LANSAA?utm_source=chatgpt.com">PhilPapers</a>)</p><p>Levy, N. (2002). Virtual child pornography: The eroticization of inequality. <em>Ethics and Information Technology, 4</em>(4), 319–323. <a href="https://doi.org/10.1023/A:1021372601566">https://doi.org/10.1023/A:1021372601566</a> (<a href="https://researchers.mq.edu.au/en/publications/virtual-child-pornography-the-eroticization-of-inequality/?utm_source=chatgpt.com">Macquarie University Researchers</a>)</p><p>Luck, M. (2009). The gamer’s dilemma: An analysis of the arguments for the moral distinction between virtual murder and virtual paedophilia. <em>Ethics and Information Technology, 11</em>(1), 31–36. <a href="https://doi.org/10.1007/s10676-008-9168-4">https://doi.org/10.1007/s10676-008-9168-4</a> (<a href="https://researchoutput.csu.edu.au/en/publications/the-gamers-dilemma-an-analysis-of-the-arguments-for-the-moral-dis?utm_source=chatgpt.com">researchoutput.csu.edu.au</a>)</p><p>Luck, M. (2022). The grave resolution to the gamer’s dilemma: An argument for a moral distinction between virtual murder and virtual child molestation. <em>Philosophia, 50</em>(3), 1287–1308. <a href="https://doi.org/10.1007/s11406-021-00455-y">https://doi.org/10.1007/s11406-021-00455-y</a> (<a href="https://philpapers.org/rec/LUCTGR-2?utm_source=chatgpt.com">PhilPapers</a>)</p><p>Madary, M., &amp; Metzinger, T. K. (2016). Real virtuality: A code of ethical conduct. Recommendations for good scientific practice and the consumers of VR-technology. <em>Frontiers in Robotics and AI, 3</em>, Article 3. <a href="https://doi.org/10.3389/frobt.2016.00003">https://doi.org/10.3389/frobt.2016.00003</a> (<a href="https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2016.00003/full?utm_source=chatgpt.com">Frontiers</a>)</p><p>McCormick, M. (2001). Is it wrong to play violent video games? <em>Ethics and Information Technology, 3</em>(4), 277–287. <a href="https://doi.org/10.1023/A:1013802119431">https://doi.org/10.1023/A:1013802119431</a> (<a href="https://link.springer.com/article/10.1023/A%3A1013802119431?utm_source=chatgpt.com">SpringerLink</a>)</p><p>Montefiore, T., &amp; Formosa, P. (2023). Crossing the fictional line: Moral graveness, the gamer’s dilemma, and the paradox of fictionally going too far. <em>Philosophy &amp; Technology, 36</em>(3), 1–21.</p><p>Murdoch, I. (1997). The idea of perfection. In <em>The sovereignty of good</em> (pp. 1–45). Routledge. (Original lecture delivered 1962; first published 1964)(<a href="https://plato.stanford.edu/entries/murdoch/?utm_source=chatgpt.com">Stanford Encyclopedia of Philosophy</a>)</p><p>Nguyen, C. T. (2020). <em>Games: Agency as art</em>. Oxford University Press.(<a href="https://academic.oup.com/book/32137?utm_source=chatgpt.com">OUP Academic</a>)</p><p>Nussbaum, M. C. (2001). <em>Upheavals of thought: The intelligence of emotions</em>. Cambridge University Press.(<a href="https://books.google.com.au/books/about/Upheavals_of_Thought.html?id=Mji-Ah10AesC&amp;utm_source=chatgpt.com">books.google.com.au</a>)</p><p>Nay, J. L., &amp; Zagal, J. P. (2017). Meaning without consequence: Virtue ethics and inconsequential choices in games. In <em>Proceedings of the 12th International Conference on the Foundations of Digital Games</em> (pp. 1–10). ACM.(<a href="https://www.researchgate.net/publication/319370603_Meaning_without_consequence_virtue_ethics_and_inconsequential_choices_in_games?utm_source=chatgpt.com">ResearchGate</a>)</p><p>Ostritsch, S. (2017). The amoralist challenge to gaming and the gamer’s moral obligation. <em>Ethics and Information Technology, 19</em>(2), 97–108. <a href="https://doi.org/10.1007/s10676-017-9420-x">https://doi.org/10.1007/s10676-017-9420-x</a> (<a href="https://philpapers.org/rec/OSTTAC-2?utm_source=chatgpt.com">PhilPapers</a>)</p><p>Patridge, S. (2010). The incorrigible social meaning of video game imagery. <em>Ethics and Information Technology, 13</em>(4), 303–312. <a href="https://doi.org/10.1007/s10676-010-9250-6">https://doi.org/10.1007/s10676-010-9250-6</a> (<a href="https://philpapers.org/rec/PATTIS?utm_source=chatgpt.com">PhilPapers</a>)</p><p>Patridge, S. L. (2013). Pornography, ethics, and video games. <em>Ethics and Information Technology, 15</em>(1), 25–34. <a href="https://doi.org/10.1007/s10676-012-9310-1">https://doi.org/10.1007/s10676-012-9310-1</a> (<a href="https://philpapers.org/rec/PATPEA-2?utm_source=chatgpt.com">PhilPapers</a>)</p><p>Scanlon, T. M. (2008). <em>Moral dimensions: Permissibility, meaning, blame</em>. Harvard University Press.(<a href="https://philpapers.org/rec/SCAMDP?utm_source=chatgpt.com">PhilPapers</a>)</p><p>Ulbricht, S. (2023). A Kantian response to the gamer’s dilemma. <em>Ethics and Information Technology, 25</em>, 1–13.(<a href="https://www.researchgate.net/publication/372234064_A_Kantian_response_to_the_Gamer%27s_Dilemma?utm_source=chatgpt.com">ResearchGate</a>)</p><p>Waldron, J. (2012). <em>The harm in hate speech</em>. Harvard University Press.(<a href="https://www.amazon.com.au/Harm-Hate-Speech-Jeremy-Waldron/dp/0674065891?utm_source=chatgpt.com">Amazon Australia</a>)</p><p>Walton, K. (1994). Morals in fiction and fictional morality. <em>Proceedings of the Aristotelian Society, 68</em>(1), 27–50.(<a href="https://philpapers.org/rec/WALMIF?utm_source=chatgpt.com">PhilPapers</a>)</p><p>Young, G. (2017). Objections to Ostritsch’s argument in “The amoralist challenge to gaming and the gamer’s moral obligation.” <em>Ethics and Information Technology, 19</em>(2), 109–117. <a href="https://doi.org/10.1007/s10676-017-9437-1">https://doi.org/10.1007/s10676-017-9437-1</a> (<a href="https://link.springer.com/article/10.1007/s10676-017-9437-1?utm_source=chatgpt.com">SpringerLink</a>)</p><p>Young, G., &amp; Whitty, M. (2010). Games without frontiers: On the moral and psychological implications of violating taboos within multi-player virtual spaces. <em>Computers in Human Behavior, 26</em>(6), 1228–1236. <a href="https://doi.org/10.1016/j.chb.2010.03.023">https://doi.org/10.1016/j.chb.2010.03.023</a> (<a href="https://dl.acm.org/doi/abs/10.1016/j.chb.2010.03.023?utm_source=chatgpt.com">ACM Digital Library</a>)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=faed268d0d5b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Ontological Gap: A Philosophical Analysis of World Models, Causal Structure, and the Limits of…]]></title>
            <link>https://medium.com/@luan.home/the-ontological-gap-a-philosophical-analysis-of-world-models-causal-structure-and-the-limits-of-470a805e9376?source=rss-67119f2f76c9------2</link>
            <guid isPermaLink="false">https://medium.com/p/470a805e9376</guid>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Ratiomachina]]></dc:creator>
            <pubDate>Thu, 20 Nov 2025 09:58:03 GMT</pubDate>
            <atom:updated>2025-11-20T09:58:03.875Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Ontological Gap: A Philosophical Analysis of World Models, Causal Structure, and the Limits of Synthetic Representation</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W8EAYyv-13aXRXaDtR1PhQ.png" /></figure><h3>1. Introduction: The Divergence of Statistical Mimicry and Causal Understanding</h3><p>The trajectory of artificial intelligence has reached a critical bifurcation point, characterized by a deepening schism between two distinct epistemological frameworks. On one side lies the commercially dominant paradigm of Large Language Models (LLMs), systems that operate primarily on the statistical distribution of discrete tokens. These models, exemplified by the GPT series, function as sophisticated engines of syntax, modeling the probability of human expression rather than the physical reality those expressions describe (LeCun, 2022; Mitchell, 2023). On the other side, a growing cohort of researchers, arguably led by Yann LeCun and the intellectual lineage of Ha and Schmidhuber (2018), advocates for “World Models” — architectures designed to endow machines with an internal representation of physical structure, dynamics, and cause-and-effect relationships (Ha &amp; Schmidhuber, 2018; LeCun, 2022).</p><p>This divergence is not merely a matter of engineering preference or architectural efficiency; it represents a profound philosophical dispute regarding the nature of intelligence, the requirements of planning, and the ontological status of machine representations. The premise of the World Model approach is that genuine intelligence requires a “simulator” of the environment — a cognitive mechanism capable of predicting the outcomes of actions before they are executed, thereby enabling counterfactual reasoning and strategic planning (LeCun, 2022). Unlike the autoregressive generation of text, which LeCun critiques as lacking a “grounded” understanding of physical reality and prone to hallucination due to its disconnect from the causal substrate of the world, world models rely on embodied data modalities such as video, spatial information, and haptic feedback to construct a latent representation of the world’s state (Kahan, 2025).</p><p>The critique of the current paradigm posits that “blind empiricism” — the reliance on massive datasets without structural priors — leads to systems that are epistemically fragile (Pearl &amp; Mackenzie, 2018). They mimic the shadows on the wall of Plato’s cave (text) without understanding the objects casting them (the physical world). The proposed remedy is an architectural shift towards systems that learn “how the world works” through observation and interaction, constructing abstract representations that are robust to noise and capable of simulating future states (LeCun, 2022; Xing et al., 2025). This article provides a rigorous philosophical analysis of the “World Model” argument. By synthesizing perspectives from structural realism, causal inference, phenomenology, and 4E cognition (Embodied, Embedded, Enacted, Extended), we examine whether the proposed architectures truly bridge the gap between statistical prediction and genuine understanding.</p><h3>2. The Architecture of Intelligibility: From Generative to Predictive Models</h3><p>To critique the philosophical status of world models, one must first dissect their technical constitution. The evolution from “model-free” reinforcement learning to “model-based” architectures represents a shift from behaviorism — learning actions based solely on reward history — to a form of cognitive constructivism where the agent builds a mental map of its domain.</p><h3>2.1 The Generative Fallacy and the JEPA Alternative</h3><p>Early iterations of world models utilized Variational Autoencoders (VAEs) combined with Recurrent Neural Networks (RNNs) to create a “dream” environment in which an agent could train (Ha &amp; Schmidhuber, 2018). These generative models operated on the principle of pixel-level reconstruction: the model attempts to predict the exact visual sensory data of the next time step given the current state. While successful in simple domains, this approach encounters a fundamental epistemological barrier in the real world: the stochasticity of the irrelevant.</p><p>Yann LeCun’s critique of this generative approach highlights a critical flaw: the physical world is replete with details that are informationally dense and unpredictable but semantically irrelevant for intelligent action (LeCun, 2022). A system attempting to model the world by predicting every pixel wastes computational resources on “nuisance variables” rather than capturing causal structures. Philosophically, this is an argument against <em>naive realism</em> in representation; an intelligent map must not be a 1:1 replica of the territory.</p><p>LeCun’s proposed solution, the <strong>Joint Embedding Predictive Architecture (JEPA)</strong>, fundamentally alters the representational ontology. Instead of predicting the raw sensory data (<em>x</em>), JEPA projects the input into an abstract representation space (<em>sx</em>​) and predicts the <em>representation</em> of the future state (<em>sy</em>​) (Assran et al., 2023; LeCun, 2022). Mathematically, this moves the prediction target from the extensional world of pixels to the intensional world of abstract features. This architecture allows the system to discard unpredictable details while retaining the predictable dynamics of relevant objects, effectively implementing a form of “transformational abstraction” (Buckner, 2018).</p><h3>2.2 Structural Realism in Latent Space</h3><p>The shift to representation-space prediction aligns the World Model approach with <strong>Structural Realism</strong> in the philosophy of science — the view that science (or in this case, the AI) does not describe the “intrinsic nature” of objects, but rather the structural relations between them (Ladyman &amp; Ross, 2007). If a world model can accurately predict the trajectory of a ball in its latent space without reconstructing the texture of the ball, it has captured the “real pattern” of the ball’s dynamics, a concept elucidated by Daniel Dennett (1991). Dennett argues that a pattern is “real” if its identification allows for information compression and prediction better than a random baseline. By this metric, the latent representations in JEPA are not merely convenient compressions; they are ontologically significant structures that map to the causal invariants of the environment.</p><p>However, this raises the “Isomorphism Problem.” As noted by Newman (1928) and discussed in modern contexts of neural representations, structural isomorphism is trivially easy to find if the mapping relations are not constrained (Shea, 2014; Frigg &amp; Votsis, 2011). The World Model proponent argues that the <em>predictive</em> requirement constrains the isomorphism: the structure of the latent space must mirror the temporal and causal structure of reality for prediction to be successful.</p><h3>2.3 Empiricism without Magic: The Defense of Deep Abstraction</h3><p>Critics often dismiss deep learning representations as “mere statistics,” but philosopher Cameron Buckner (2018) offers a robust defense of these architectures as implementing a sophisticated form of empiricism. Buckner argues that Deep Convolutional Neural Networks (DCNNs) and by extension, hierarchical architectures like H-JEPA, solve the longstanding rationalist complaint against empiricism: the failure to explain <em>how</em> abstraction works. Through “transformational abstraction,” these systems iteratively convert sensory-based representations into new formats that are increasingly tolerant to “nuisance variation” (e.g., changes in lighting, pose, or background) (Buckner, 2018).</p><h3>3. The Ladder of Causation: Observation vs. Intervention</h3><p>The central promise of world models is the ability to “reason about cause and effect” and “simulate future outcomes” (LeCun, 2022). However, the reliance on video pretraining as a primary data source invokes the limitations described in Judea Pearl’s “Ladder of Causation” (Pearl &amp; Mackenzie, 2018).</p><h3>3.1 The Identifiability Problem in Passive Observation</h3><p>Pearl distinguishes between three levels of cognitive ability:</p><ol><li><strong>Association:</strong> Seeing/Observing (correlation).</li><li><strong>Intervention:</strong> Doing/Acting (causality).</li><li><strong>Counterfactuals:</strong> Imagining (what if?).</li></ol><p>Current Large Language Models reside firmly on the first rung, effectively calculating conditional probabilities based on massive textual correlations (Pearl &amp; Mackenzie, 2018). World model proponents argue that by training on video (embodied data), AI can climb to the second and third rungs. The argument posits that video contains implicit causal physics — gravity, collision, object permanence — that the model can internalize (Assran et al., 2023).</p><p>However, a critical philosophical and mathematical hurdle remains: <strong>Identifiability</strong>. Causal structure is often underdetermined by observational data alone (Bareinboim et al., 2022; Pearl, 2009). Without the ability to intervene (<em>do</em>(<em>x</em>)), an observer cannot rigorously distinguish between causation (<em>A</em>→<em>B</em>) and a hidden confounder (<em>C</em> causing both <em>A</em> and <em>B</em>). If a world model is pretrained primarily on passive video, it is learning a “spectator’s physics.” It observes that “when the handle turns, the door opens,” but without the proprioceptive feedback of <em>turning</em> the handle, the causal link is inferred, not grounded.</p><h3>3.2 The Do-Calculus in Latent Dynamics</h3><p>LeCun’s architecture attempts to circumvent this by explicitly including an action vector <em>a</em> in the predictive function (LeCun, 2022). In theory, this formalizes Pearl’s “do-operator.” The model learns to predict how the state changes <em>given</em> an action. However, if the model is trained in a passive mode (watching others act), the action variable is just another observation. To truly learn interventional causality, the agent must be able to select actions independent of the current state’s confounders (Pearl, 2009). This distinction is subtle but vital: watching a cooking video is different from cooking. The former builds a model of <em>sequences</em>; the latter builds a model of <em>consequences</em>.</p><h3>4. Embodiment and the Phenomenology of “Being-in-the-World”</h3><p>The critique of world models deepens when viewed through the lens of phenomenology and 4E cognition. The “embodied data” relied upon by these models — video, spatial coordinates, and touch sensors — is fundamentally different from the “embodiment” described by philosophers like Merleau-Ponty and Hubert Dreyfus.</p><h3>4.1 The Spectator vs. The Agent</h3><p>Hubert Dreyfus famously critiqued “Good Old-Fashioned AI” (GOFAI) for assuming that intelligence consists of manipulating formal symbols representing the world (Dreyfus, 1972/1992). While modern deep learning and JEPA architectures abandon explicit symbols for continuous vectors, they still adhere to the “modeling” paradigm: the idea that the mind is a mirror of reality.</p><p>Dreyfus, drawing on Heidegger, argued that human intelligence is primarily “coping” — a non-representational, skilful engagement with the world (Dreyfus, 2007). We do not navigate a room by consulting an internal 3D map; we navigate because the room affords walking. The door handle is not a geometrical object with properties; it is “ready-to-hand” (<em>Zuhanden</em>), defined by its use (Heidegger, 1927/1962). World models, conversely, treat the world as “present-at-hand” (<em>Vorhanden</em>) — a collection of objects and states to be computed and predicted.</p><h3>4.2 The Symbol Grounding Problem in Vector Space</h3><p>The “Symbol Grounding Problem,” articulated by Stevan Harnad (1990), asks how symbols in a system acquire meaning intrinsic to the system rather than just referring to other symbols. Proponents argue that multimodal world models solve this by grounding text in video and sensory data (Millière, 2024).</p><p>However, a rigorous analysis suggests this is “sensor grounding,” not true symbol grounding. Mapping a text token to a visual tensor is simply mapping one digital code to another. Proper grounding, as argued in enactive cognitive science, arises from the <em>consequences</em> of action — the resistance the world offers to the agent’s will (Varela, Thompson, &amp; Rosch, 1991). In virtual environments or video pretraining, the “resistance” is simulated logic or pixel arrangements, not physical law. Consequently, the representations in JEPA are “grounded” only in the sense that they are consistent with the statistical structure of the input data, lacking the “existential grounding” that comes from the vulnerability of a biological organism (Deng et al., 2025).</p><h3>5. Active Inference vs. The Configurator: The Locus of Agency</h3><p>A critical comparison arises between LeCun’s architecture and Karl Friston’s <strong>Active Inference</strong> framework. Both posit that the brain minimizes prediction error (or free energy), yet their architectural implementation of agency differs radically.</p><h3>5.1 The Configurator as Homunculus</h3><p>LeCun’s architecture features a “Configurator” module — a central controller that sets subgoals and modulates the parameters of the world model and cost module (LeCun, 2022). Philosophically, this approach risks reintroducing a homuncular regress: who configures the configurator? If the configurator is merely trained via gradients to minimize a cost, it acts as a reactive component rather than a seat of agency.</p><h3>5.2 Active Inference and Autopoiesis</h3><p>In contrast, Active Inference views the agent and the world model as co-emergent. The agent does not just “have” a model; it <em>is</em> a model of its niche (Friston, 2013). Under the Free Energy Principle, the goal is not just to predict the future (as in JEPA), but to <em>act</em> to make the future conform to the agent’s viable states (homeostasis) (Kugele, 2023). The critique here is that LeCun’s approach prioritizes <em>epistemic accuracy</em> (knowing the world), whereas biological intelligence prioritizes <em>adaptive success</em> (surviving the world).</p><h3>6. The Commercial Temporality of Cognitive Architectures</h3><p>The observation that world models “require new cognitive architectures that are slower to commercialize” highlights a friction between economic incentives and ontological necessity (Kahan, 2025).</p><h3>6.1 System 1 vs. System 2 Scaling</h3><p>The current commercial explosion of generative AI is driven by the scalability of “System 1” thinking — fast, intuitive, pattern-matching responses generated by LLMs (LeCun, 2022; Sequoia Capital, 2024). These models scale with data volume and compute. World models, representing “System 2” (deliberate reasoning, planning, simulation), require the integration of <em>time</em> and <em>causality</em> into the training loop (LeCun, 2022).</p><h3>6.2 The Economic “Frame Problem”</h3><p>The “slowness” to commercialize stems from the difficulty of verifying the “truth” of a world model. An LLM hallucination is a syntax error or a factual error, often acceptable in creative tasks. A world model hallucination is a violation of physics, which is catastrophic for embodied agents. To build a commercially viable world model, one must solve the “Frame Problem” in engineering terms: ensuring the model understands which physical laws are invariant across contexts without manually encoding them (McCarthy &amp; Hayes, 1969; Shanahan, 2016).</p><h3>7. Virtual Realism and the Status of Simulated Worlds</h3><p>If World Models succeed in creating high-fidelity internal simulators, what is the ontological status of the objects they simulate? David Chalmers’ (2017) “Virtual Realism” argues that virtual objects are <em>real</em> objects with digital properties. If a World Model simulates a cup falling, and that simulation is causally responsible for the robot catching the cup, the simulated cup has causal power, satisfying a key criterion of reality (Chalmers, 2022).</p><p>However, Thomas Metzinger (2018) warns of the “Phenomenological Gap.” Even if the machine’s simulation is structurally real, it lacks the <em>phenomenal transparency</em> that characterizes human experience. A machine that manipulates latent variables is manipulating opaque data structures. Unless the machine has a “self-model” that is transparently identified with the system itself, it remains a zombie simulator — competent but unconscious (Metzinger, 2018).</p><h3>8. Conclusion</h3><p>The “World Model” approach represents a necessary correction to the “blind empiricism” of large language models. By seeking to construct internal representations of physical dynamics, architectures like JEPA move closer to the “structural realism” that characterizes scientific inquiry. However, without active intervention during training, these models struggle to distinguish correlation from causation (The Identifiability Problem). Furthermore, “embodied data” without biological vulnerability constitutes a “digital embodiment” that simulates the optics of agency without the ontology of “being-in-the-world.” Ultimately, while World Models may succeed in creating powerful <em>simulators</em>, the path to true autonomous intelligence requires closing the loop between the model and the world, transforming the machine from a predictor of pixels to an agent of its own survival.</p><h3>References</h3><p>Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y., &amp; Ballas, N. (2023). Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture. <em>Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)</em>.</p><p>Bareinboim, E., Correa, J., Ibeling, D., &amp; Icard, T. (2022). On Pearl’s Hierarchy and the Foundations of Causal Inference. In <em>Probabilistic and Causal Inference: The Works of Judea Pearl</em> (pp. 507–556). ACM.</p><p>Buckner, C. (2018). Empiricism without Magic: Transformational Abstraction in Deep Convolutional Neural Networks. <em>Synthese</em>, 195(12), 5339–5372.</p><p>Chalmers, D. J. (2017). The Virtual and the Real. <em>Disputatio</em>, 9(46), 309–352.</p><p>Chalmers, D. J. (2022). <em>Reality+: Virtual Worlds and the Problems of Philosophy</em>. W. W. Norton &amp; Company.</p><p>Deng, M., Xing, E., &amp; Hu, Z. (2025). Critiques of World Models. <em>arXiv preprint arXiv:2507.05169</em>.</p><p>Dennett, D. C. (1991). Real Patterns. <em>The Journal of Philosophy</em>, 88(1), 27–51.</p><p>Dreyfus, H. L. (1992). <em>What Computers Still Can’t Do: A Critique of Artificial Reason</em>. MIT Press. (Original work published 1972).</p><p>Dreyfus, H. L. (2007). Why Heideggerian AI Failed and How Fixing it Would Require Making it More Heideggerian. <em>Philosophical Psychology</em>, 20(2), 247–268.</p><p>Friston, K. (2013). Life as we know it. <em>Journal of the Royal Society Interface</em>, 10(86).</p><p>Frigg, R., &amp; Votsis, I. (2011). Everything you always wanted to know about structural realism but were afraid to ask. <em>European Journal for Philosophy of Science</em>, 1, 227–276.</p><p>Ha, D., &amp; Schmidhuber, J. (2018). World Models. <em>NeurIPS 2018</em>.</p><p>Harnad, S. (1990). The Symbol Grounding Problem. <em>Physica D: Nonlinear Phenomena</em>, 42(1–3), 335–346.</p><p>Heidegger, M. (1962). <em>Being and Time</em> (J. Macquarrie &amp; E. Robinson, Trans.). Harper &amp; Row. (Original work published 1927).</p><p>Kahan, R. (2025, February 18). AI godfather warns language models are a ‘dead end,’ says world models are the future. <em>Ynet News</em>.</p><p>Kugele, S. C. (2023). <em>Embodied, Simulation-Based Cognition: A Hybrid Approach</em> (Doctoral dissertation). University of Memphis.</p><p>Ladyman, J., &amp; Ross, D. (2007). <em>Every Thing Must Go: Metaphysics Naturalized</em>. Oxford University Press.</p><p>LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. <em>OpenReview</em>.</p><p>McCarthy, J., &amp; Hayes, P. J. (1969). Some Philosophical Problems from the Standpoint of Artificial Intelligence. <em>Machine Intelligence</em>, 4, 463–502.</p><p>Metzinger, T. K. (2018). Why is Virtual Reality Interesting for Philosophers? <em>Frontiers in Robotics and AI</em>, 5, 101.</p><p>Millière, R. (2024). World Models in AI and Philosophy. <em>Synthese</em>.</p><p>Mitchell, M. (2023). The debate over understanding in AI’s large language models. <em>Proceedings of the National Academy of Sciences</em>.</p><p>Pearl, J. (2009). <em>Causality: Models, Reasoning, and Inference</em> (2nd ed.). Cambridge University Press.</p><p>Pearl, J., &amp; Mackenzie, D. (2018). <em>The Book of Why: The New Science of Cause and Effect</em>. Basic Books.</p><p>Sequoia Capital. (2024, September). Generative AI’s Act o1.</p><p>Shanahan, M. (2016). The Frame Problem. <em>The Stanford Encyclopedia of Philosophy</em>.</p><p>Shea, N. (2014). Exploitable Isomorphism and Structural Representation. <em>Proceedings of the Aristotelian Society</em>, 114(2), 123–144.</p><p>Varela, F. J., Thompson, E., &amp; Rosch, E. (1991). <em>The Embodied Mind: Cognitive Science and Human Experience</em>. MIT Press.</p><p>Xing, E., Deng, M., Hou, J., &amp; Hu, Z. (2025). Critiques of World Models. <em>arXiv preprint arXiv:2507.05169</em>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=470a805e9376" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The World in the Model: A Critique of AI's “Internal Representations”]]></title>
            <link>https://medium.com/@luan.home/the-world-in-the-model-a-critique-of-ais-internal-representations-5913c503337b?source=rss-67119f2f76c9------2</link>
            <guid isPermaLink="false">https://medium.com/p/5913c503337b</guid>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[large-world-model]]></category>
            <dc:creator><![CDATA[Ratiomachina]]></dc:creator>
            <pubDate>Tue, 18 Nov 2025 09:02:54 GMT</pubDate>
            <atom:updated>2025-11-18T09:02:54.504Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7CY3-lFwFB98KeRtsXwa9g.png" /></figure><h3>Abstract</h3><p>The emergence of large-scale “world models” in artificial intelligence (AI) has prompted claims that AI is approaching a human-like capacity for planning and understanding. Proponents now frame these systems as nascent <strong>“world simulators” (OpenAI, 2024)</strong>, arguing that by processing “embodied data” (e.g., video, touch) to create internal representations of physical dynamics, these systems can predict outcomes “similar to how humans and animals plan.”</p><p>I argue that such claims rest on a profound <strong>category error</strong>, conflating <strong>functional simulation with genuine understanding</strong>.</p><p>Drawing on contemporary frameworks from philosophy of mind and language, this critique deconstructs the analogy at three levels:</p><ol><li><strong>Meaning:</strong> The AI’s “internal representation” is a mastery of statistical <strong>form</strong>, but it lacks access to <strong>meaning</strong> (Bender &amp; Koller, 2020). It is a “Stochastic Parrot” (Bender et al., 2021) for physics, not a semantic agent.</li><li><strong>Embodiment:</strong> The model <em>processes</em> third-person data; it is not an <em>enactive</em> agent (Noë, 2009). It lacks the first-person, <em>affective</em> grounding of a biological, homeostatic organism (Seth, 2021), for whom outcomes <em>matter</em>.</li><li><strong>Causality:</strong> The model excels at correlation and intervention, but it does not achieve true causal <em>understanding</em>, which requires access to counterfactual reasoning (Pearl, 2018).</li></ol><p>I conclude that while “world simulators” are powerful predictive tools, the “similarity” to human cognition is merely superficial. The system remains a <em>formal</em> engine that <em>models</em> a dataset, not a semantic subject that <em>understands</em> a world.</p><h3>1. Introduction: The Limits of the Analogy</h3><p>The pursuit of artificial general intelligence (AGI) has recently coalesced around the concept of “world models.” While this line of research has foundational papers (e.g., Ha &amp; Schmidhuber, 2018), the philosophical claims undergirding it have become significantly more ambitious. The latest generation of large-scale generative systems, such as <strong>Google’s Genie (Reed et al., 2024)</strong> and <strong>OpenAI’s Sora (OpenAI, 2024)</strong>, are no longer just discussed as predictive tools. They are explicitly framed by their creators as nascent <strong>“world simulators”</strong> (OpenAI, 2024) or “foundation world models” (Reed et al., 2024).</p><p>Proponents argue that by training on vast datasets of video and interaction, these models build “internal representations” of physical dynamics and causal rules. This, they claim, allows the AI to “predict the outcomes of actions” and plan behaviour <em>“similar to how humans and animals plan.”</em></p><p>As an engineering thesis, this approach is powerful. Creating high-fidelity predictive simulations is a clear benchmark of progress in autonomous control. The philosophical claims embedded within this technical description, however, are far more ambitious and far less secure.</p><p>This article argues that the language of “internal representation” and “similarity to human planning” masks a profound category error. It <strong>conflates the successful simulation of a process with the possession of the cognitive states</strong> that accompany that process in humans.</p><p>The AI’s “world model” is a sophisticated, high-dimensional statistical map. The human “world model” is a subjective, intentional, and phenomenologically-grounded <em>experience</em>. This paper will deconstruct the “similarity” claim by demonstrating that the AI model lacks three foundational components of human cognition: <strong>semantic meaning</strong>, <strong>enactive embodiment</strong>, and <strong>counterfactual causal reasoning</strong>.</p><h3>2. The Problem of Representation: Syntax vs. Semantics</h3><p>The argument for world models hinges on the term “internal representation.” In a human, a representation is <em>about</em> something (i.e., it has intentionality). This “aboutness” is a <em>semantic</em> property; the thought <em>means</em> something.</p><p>The AI’s “representation,” by contrast, is a complex mathematical state derived from statistical correlations in a dataset. This issue is a direct parallel to the critique of Large Language Models (LLMs). <strong>Bender and Koller (2020)</strong> argue that models trained <em>only</em> on <strong>form</strong> (the statistical patterns of language) can never, by definition, learn <strong>meaning</strong> (which is grounded in communicative intent and a shared external world). The “Stochastic Parrot” (Bender et al., 2021) can mimic human language flawlessly without <em>meaning</em> a single word.</p><p>This critique extends perfectly to “world simulators.”</p><ul><li>By processing petabytes of video, a system like Sora learns the <strong>“form” of physics</strong> — the statistical regularities of “what pixels are likely to follow what pixels.” It can predict that a glass striking a floor will be followed by pixels representing shards.</li><li>It does not, however, learn the <strong>“meaning”</strong> of this event. It lacks access to the concepts of <em>fragility</em>, <em>gravity</em>, and <em>physicality</em>. It is a Stochastic Parrot for physics.</li></ul><p>Proponents might argue that “embodied data” (such as video and touch) <em>is</em> the grounding that language models lack. But this pushes the problem back. The model isn’t “embodying” anything; it is simply processing a <em>different data modality</em>. The video data is just another stream of formal tokens. Without a genuine, <em>enactive</em> grounding in the world (which we will explore in Section 3), the model remains a syntactic engine, possessing only <em>derived intentionality</em> (Searle, 1983) — it is meaningful <em>to us</em>, not to itself.</p><h3>3. The Problem of Embodiment: Data vs. Lived Experience</h3><p>The world model argument correctly identifies embodiment as crucial to intelligence. Its mistake is confusing the <em>processing of embodied data</em> with the <em>fact of lived, phenomenological embodiment</em>.</p><p>This critique rests on two pillars of contemporary philosophy of mind:</p><ol><li><strong>Enactive Cognition (The “Brain in a Vat” Fallacy):</strong> Philosopher <strong>Alva Noë (2009)</strong> argues that consciousness and understanding are not <em>processes</em> that a brain (or a model) <em>runs</em>. They are an <em>activity</em> that a whole, living, embodied organism <em>does</em> by actively engaging with its environment. Understanding is an <em>enactive</em> skill, not a passive “internal model.”</li><li>The AI “world simulator” is the quintessential “brain in a vat.” It is a disembodied intelligence receiving a <em>data feed</em> of a world it cannot act in. A model like <strong>Genie (Reed et al., 2024)</strong>, trained on 2D platformer videos, isn’t <em>being</em> a body; it’s modelling the <em>data from</em> one. This is the difference between <em>having</em> a world and <em>modelling</em> a dataset.</li><li><strong>Biological Affect (The “Mattering” Problem):</strong> Why is human planning different from AI “planning”? Because, for us, the outcomes <em>matter</em>. <strong>Anil Seth (2021)</strong> provides a modern, biological basis for this. He argues that our core subjective consciousness is not a high-level intellectual function but is deeply rooted in our biological nature as <em>living organisms</em>.</li><li>Our “world model” is not just for seeing, but for <em>surviving</em>. It is governed by <strong>homeostasis</strong> — the brain’s constant predictive modelling of our internal bodily state (our “feeling of being alive”) to ensure our survival. This is the root of <em>affect</em>, <em>emotion</em>, and <em>concern</em>.</li><li>The AI “world simulator” is a purely <em>informational</em> system. It has no body to protect, no homeostatic state to maintain. Its “planning” is therefore “disinterested.” It lacks the <em>affective stake</em> in the outcome that is the foundation of all human (and animal) planning.</li></ol><h3>4. The Problem of Causality: Correlation vs. Understanding</h3><p>The “world simulator” claim rests on the idea that models understand “cause-and-effect relationships” to “predict the outcomes of actions.” This claim misidentifies the nature of causal understanding, mistaking <strong>probabilistic correlation for genuine causal reasoning.</strong></p><p><strong>OpenAI (2024)</strong> itself notes in its technical report on Sora that the model can struggle with “simulating the physics of a complex scene” or “understanding specific instances of cause and effect” (e.g., a person eating a cookie that does not subsequently show a bite mark).</p><p>This admission is philosophically revealing: the model has mastered <strong>correlation</strong>, not <strong>causation</strong>. It “learns” that state A (a hand releasing a ball) is <em>consistently followed by</em> state B (the ball on the floor). It does not, however, <em>understand</em> the “why” — the concepts of <em>gravity</em>, <em>force</em>, or <em>necessity</em> that govern this interaction.</p><p>The most rigorous framework for this critique comes from <strong>Judea Pearl’s (2018) “Ladder of Causation.”</strong> Pearl defines three distinct levels of cognitive ability:</p><ol><li><strong>Rung 1: Association (Seeing).</strong> This is <em>observing</em> correlations. (e.g., “I see that people who have a cold often sneeze.”) All deep learning, including world models, excels at this.</li><li><strong>Rung 2: Intervention (Doing).</strong> This is <em>predicting</em> the effects of a deliberate action. (e.g., “What will happen <em>if I</em> drop the ball?”) This is precisely what AI world models are designed to do. A model like Sora can predict <em>that</em> a glass will shatter (Rung 2). Proponents mistake success at this rung for true understanding.</li><li><strong>Rung 3: Counterfactuals (Imagining).</strong> This is the ability to <em>reason about</em> alternate possibilities and “why” things happened as they did. (e.g., “What <em>would have happened</em> if the ball had been made of helium?” or “The glass broke. <em>Would it have</em> broken if I had dropped it on the carpet?”)</li></ol><p>According to Pearl, Rung 3 is the hallmark of genuine causal understanding. Human (and many animal) planners live on this rung. We plan by <em>imagining</em> possibilities that have not occurred.</p><p>The AI “world model” is “stuck” between Rungs 1 and 2. The very existence of new, specialised research into <em>explicitly</em> causal, counterfactual video generation (e.g., von Kügelgen et al., 2024) demonstrates that current “world simulators” do not possess this (Rung 3) capability.</p><h3>5. Conclusion: A Category Error in Similarity</h3><p>The “world simulator” approach (OpenAI, 2024; Reed et al., 2024) is a powerful engineering achievement. It creates predictive tools of immense value. But the philosophical claim that this architecture allows an AI to plan “similar to how humans and animals plan” is untenable.</p><p>The “similarity” is one of <strong>functional output only</strong>, and this analogy is achieved by systematically ignoring what makes human cognition unique:</p><ul><li><strong>Humans:</strong> Experience the world as an <em>enactive, affective</em> agent (Noë, 2009; Seth, 2021) → Form <em>meaningful, intentional</em> states grounded in survival → Use a rich, <em>counterfactual</em> understanding of causality (Pearl, 2018) → Plan an action.</li><li><strong>AI World Model:</strong> Processes <em>formal patterns</em> (Bender &amp; Koller, 2020) → Calculates <em>probabilistic correlations</em> → “Selects” an optimal output path.</li></ul><p>To call these processes “similar” is to commit a category error. A system like Sora or Genie does not “know” what its model represents. It does not “feel” the data it embodies. It does not “understand” the causes it predicts. The “world simulator” is a remarkable system, but it is not a mind. It does not <em>have</em> a world; it <em>models</em> a dataset.</p><h3>Works Cited</h3><ul><li>Bender, E. M., &amp; Koller, A. (2020). “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data.” <em>Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL)</em>.</li><li>Bender, E. M., Gebru, T., McMillan-Major, A., &amp; Shmitchell, S. (2021). “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” <em>Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT)</em>.</li><li>Ha, D., &amp; Schmidhuber, J. (2018). “World Models.” <em>arXiv:1803.10122</em>.</li><li>Hafner, D., et al. (2023). “Mastering Diverse Domains through World Models.” <em>arXiv:2301.04104</em>.</li><li>Noë, A. (2009). <em>Out of Our Heads: Why You Are Not Your Brain, and Other Lessons from the Biology of Consciousness</em>. Hill and Wang.</li><li>OpenAI. (2024). “Video generation models as world simulators.” <em>OpenAI Technical Report</em>. Retrieved from [OpenAI website].</li><li>Pearl, J. (2018). <em>The Book of Why: The New Science of Cause and Effect</em>. Basic Books.</li><li>Reed, S., et al. (2024). “Genie: A Generative Interactive Environment.” <em>arXiv:2402.15391</em>.</li><li>Searle, J. R. (1983). <em>Intentionality: An Essay in the Philosophy of Mind</em>. Cambridge University Press.</li><li>Seth, A. (2021). <em>Being You: A New Science of Consciousness</em>. Dutton.</li><li>von Kügelgen, J., et al. (2024). “Causal Re-enactment: What-if Video Generation for Causal Understanding.” <em>arXiv:2404.13324</em>.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5913c503337b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Are AI Models Converging on Reality? Bergson, Plato, and the Battle over Images]]></title>
            <link>https://medium.com/@luan.home/are-ai-models-converging-on-reality-bergson-plato-and-the-battle-over-images-418c164135e8?source=rss-67119f2f76c9------2</link>
            <guid isPermaLink="false">https://medium.com/p/418c164135e8</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[philosophy]]></category>
            <dc:creator><![CDATA[Ratiomachina]]></dc:creator>
            <pubDate>Tue, 30 Sep 2025 10:11:43 GMT</pubDate>
            <atom:updated>2025-09-30T10:11:43.426Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*bxC-7Cnoe5ne7kIR" /></figure><p>When we scroll through social media or query a chatbot, it&#39;s easy to think of AI systems as machines of representation — models that digest data and spit out structured forms. But a new wave of AI research argues something bolder: that neural networks across domains are <strong>converging toward a shared statistical model of reality</strong>. In other words, they&#39;re not just representing data; they&#39;re edging closer to a Platonic ideal. (<a href="https://arxiv.org/pdf/2405.07987">https://arxiv.org/pdf/2405.07987</a>)</p><p>Philosophers may feel a twinge of déjà vu here. Plato&#39;s allegory of the cave haunts this vision: we live among shadows, while truth exists in a higher realm of pure forms. Yet another voice from philosophy — Henri Bergson — whispers a different story. For Bergson, the world is not made of representations or hidden forms. It is already composed of <strong>images</strong>, authentic and external to us, overflowing the neat boundaries of cognition.</p><p>So which vision should guide our thinking about AI?</p><h3>Bergson&#39;s Radical Notion of Images</h3><p>In <em>Matter and Memory</em> (1896), Bergson shattered the representationalist frame. Perception, he argued, is not the creation of inner pictures. Instead, it is a <strong>selection from a field of images that already exist in the world</strong>.</p><blockquote>“The universe is made up of images. My perception of the universe is my selection from among these images.”<em> — Bergson</em></blockquote><p>This means that images are not internal mental constructs but <strong>ontological realities</strong>. They exist in between matter and mind, irreducible to either. To perceive is to participate in this flux of images, not to encode it.</p><p>For Bergson, then, representation is always too static, too impoverished. It freezes, becoming a snapshot, when reality itself is movement and duration.</p><h3>The AI Hypothesis: Platonic Convergence</h3><p>The paper <em>&quot;The Platonic Representation Hypothesis&quot;</em> (Huh et al., 2024) argues that deep learning models are converging in their internal space structures.</p><p>Vision models, language models, and even multimodal systems are increasingly aligned: they measure distances between data points in strikingly similar ways.</p><p>The authors hypothesise that this trend points toward a <strong>platonic representation</strong>: a shared statistical model of the underlying reality (denoted Z in their diagrams). Training data are like shadows on Plato&#39;s cave wall, but as models scale, they recover ever better approximations of the hidden structure of the world.</p><p>In short, AI research suggests that convergence is a form of epistemic progress — that different architectures and objectives are converging on the same truth.</p><h3>Where Bergson Pushes Back</h3><p>Here lies the philosophical tension.</p><ul><li><strong>Plato &amp; AI:</strong> Reality is a hidden order behind appearances. AI representations are shadows of shadows, approximating the ideal form.</li><li><strong>Bergson:</strong> Reality <em>is not behind appearances</em>. It is the very play of images in which we live and act. Representation is always a reduction of this movement.</li></ul><p>So when AI researchers claim convergence on &quot;reality,&quot; Bergson would ask: reality in what sense? Are we uncovering the living flux of images, or are we compressing it into statistical abstractions?</p><p>Convergence may signal not metaphysical access but <strong>epistemic economy</strong>: different models, constrained by optimisation and task demands, are funnelled toward similar vector spaces. That is a form of practical alignment — but it is not the same as touching reality itself.</p><h3>Why This Matters</h3><p>If we follow Plato, we might think AI is approaching a universal truth of the world. If we follow Bergson, we realise that what AI achieves is more modest: a <strong>filter on reality&#39;s images</strong>, tailored to tasks, data, and performance metrics.</p><p>This shift matters for how we evaluate AI.</p><ul><li>Do we treat AI outputs as if they mirror the world?</li><li>Or do we see them as selections that reveal some aspects of reality while obscuring others?</li></ul><p>Bergson helps us resist the temptation to mistake representation for ontology. AI models may align, but alignment is not revelation.</p><h3>Closing Reflection</h3><p>The real lesson is that AI prompts us to revisit an old philosophical riddle: is knowledge about <em>capturing reality</em>, or about <em>choosing which slice of reality to act upon</em>?</p><p>The Platonic impulse imagines a march toward truth. The Bergsonian impulse reminds us that the world already brims with images, irreducible to any one representation.</p><p>In the end, maybe AI doesn&#39;t converge on reality itself — but on the kinds of representations that we, as image-selecting beings, find most useful.</p><blockquote>And that, Bergson might say, tells us more about us than about reality.</blockquote><h3>References</h3><p>Bergson, H. (1991). <em>Matter and memory</em> (N. M. Paul &amp; W. S. Palmer, Trans.). Zone Books. (Original work published 1896)</p><p>Huh, M., Cheung, B., Wang, T., &amp; Isola, P. (2024). <em>The Platonic Representation Hypothesis</em>. Proceedings of the 41st International Conference on Machine Learning (PMLR 235). <a href="https://phillipi.github.io/prh/">https://phillipi.github.io/prh/</a></p><p>Plato. (1997). <em>Republic</em> (G. M. A. Grube, Trans., rev. C. D. C. Reeve). Hackett. (Original work published ca. 375 BCE)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=418c164135e8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Does AI Produce Knowledge? A Contemporary Epistemology of Artificial Systems]]></title>
            <link>https://medium.com/@luan.home/does-ai-produce-knowledge-a-contemporary-epistemology-of-artificial-systems-b58d6c2784a4?source=rss-67119f2f76c9------2</link>
            <guid isPermaLink="false">https://medium.com/p/b58d6c2784a4</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[human-ai-collaboration]]></category>
            <category><![CDATA[philosophy]]></category>
            <dc:creator><![CDATA[Ratiomachina]]></dc:creator>
            <pubDate>Wed, 10 Sep 2025 09:39:53 GMT</pubDate>
            <atom:updated>2025-09-10T09:39:53.146Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*S-oJlor6QhEJaTid" /></figure><h3>Introduction</h3><p>As AI rapidly infuses products and services across every sector — from finance and law to healthcare and education — it is reshaping the very conditions of our <strong>epistemic enterprise</strong>. Human societies depend on knowledge: it grounds our decisions, secures trust in institutions, and enables progress. But with artificial systems now generating outputs that look like knowledge, we face a profound question: <strong>are these outputs genuinely knowledge, or merely the appearance of it?</strong> Understanding the nature of AI-produced information is not just an academic exercise; it is essential for ensuring that our reliance on artificial agents strengthens, rather than undermines, the epistemic foundations of society.</p><h3>TL;DR</h3><p>AI systems do not <em>know</em> in the human, doxastic sense (no beliefs, no accountability), but they can still be <strong>sources of knowledge for us</strong> in three ways:</p><ol><li>As <strong>epistemic instruments</strong> (like simulations or lab devices) that extend our cognitive capacities;</li><li>by delivering <strong>artificial testimony</strong> whose outputs can sometimes justify our beliefs under reliability-first conditions; and</li><li>by scaffolding <strong>understanding</strong> when we can map model internals to mechanisms (not just correlations).</li></ol><p><strong>The upshot</strong>: treat AI primarily as <strong>cognitive artifacts</strong> and <strong>computationally reliable instruments</strong>, then add carefully delimited norms for <strong>machine testimony</strong>. That&#39;s enough to ground warranted uptake without anthropomorphising.</p><h3>1) Start with categories: knowledge in, by, and through AI</h3><p><strong>Knowledge <em>in</em> AI</strong> (whether models <em>themselves</em> know) is not required for <strong>knowledge <em>by</em> humans <em>through</em> AI</strong>. The philosophy of simulation shows that computational methods can be epistemically novel even if they aren&#39;t believers; the novelty lies in the <strong>method</strong> and what it makes knowable to us. Humphreys argued that computational science introduces new epistemic issues (e.g., opacity) and shifts humans from the centre of inquiry; this supports a non-anthropocentric frame for AI as knowledge-producing instrumentation.</p><p>In short, AI can be <strong>part of the pipeline that produces human knowledge,</strong> even if it isn&#39;t a knower. That&#39;s squarely consistent with work on <strong>cognitive artifacts</strong> and the <strong>extended mind</strong> — tools that participate functionally in cognition and inquiry.</p><h3>2) Epistemic opacity and why transparency isn&#39;t the master key</h3><p>Modern AI (and complex simulations) are often <strong>essentially opaque</strong> — no single agent can survey all epistemically relevant steps. If we demanded complete transparency to justify belief, we would forfeit much of modern computational knowledge. A better approach is <strong>computational reliabilism, which</strong> justifies uptake by demonstrating that the method reliably produces true (or accurate) outputs through design, validation, and error-monitoring regimes. This is now a leading view in the epistemology of algorithms.</p><p><strong>Practical takeaway:</strong> ask less &quot;Do we interpret every weight?&quot; and more &quot;Is the overall process reliably truth-conducive (benchmarks, calibration, adversarial testing, uncertainty control)?”</p><h3>3) Artificial testimony: Can we learn from what models say?</h3><p>Recent social-epistemology work treats LLM outputs as a form of <strong>artificial testimony</strong>. On this view, recipients can acquire knowledge from AI if the <em>system</em> manifests stable propensities akin to veracity and caution (e.g., calibrated uncertainty, consistency checks), even though it lacks human intentions. This reframes chat outputs as <em>testimony-like signals, whose justificatory force depends on engineered reliability properties, rather than</em> speaker sincerity. Debates here are active and high-level: some argue LLM testimony can transmit knowledge; others deny AI can ever be an <strong>epistemic expert</strong> or authority.</p><blockquote>Where I land: <strong>artificial testimony can justify belief only when embedded in institutional procedures that proxy human virtues</strong> — e.g., documented data provenance, out-of-distribution guards, model auditing, and routinised challenge mechanisms.</blockquote><h3>4) From prediction to understanding</h3><p>Prediction isn&#39;t yet understanding. But there&#39;s a promising path: <strong>mechanistic interpretability</strong> aims to connect model internals to algorithmic &quot;parts and processes&quot; that explain outputs. When successful, it upgrades AI from a black-box oracle to an <strong>explanatory instrument</strong>, enabling <em>understanding-why</em>, not just <em>knowing-that</em>. This mirrors how we achieve understanding from simulations: by relating computational structure to domain mechanisms.</p><p><strong>Policy-relevant nuance:</strong> when MI fails or is premature, require <em>epistemic surrogates</em> for understanding (ablation-based robustness, counterfactual testing, external validity studies). Those can still provide knowledge that, under reliabilism, can be gained through deeper mechanistic work.</p><h3>5) Guardrails: when AI degrades, rather than upgrades, our epistemic position</h3><p>Three contemporary hazards matter for whether AI <em>produces</em> knowledge or <strong>produces the look of knowledge</strong>:</p><ul><li><strong>Hallucination and overconfident misrepresentation</strong> — LLMs can generate fluent falsehoods; recent empirical work shows risk of oversimplification in scientific summarisation. Institutional checks are needed (grounding, domain-specific retrieval, uncertainty displays).</li><li><strong>Opacity-enabled deference traps</strong> — unchecked social narratives (e.g., &quot;the model is objective&quot;) can induce unwarranted <em>epistemic deference</em>, undermining critical agency (Vallor&#39;s warning about yielding judgment to machines).</li><li><strong>Information-ecosystem feedback</strong> — AI systems increasingly train on their own outputs, risking <strong>epistemic lock-in</strong> and &quot;anachronism&quot; (knowledge that recursively amplifies its own errors and omissions). This is an emerging, nontrivial collective risk.</li></ul><h3>6) A practical standard for &quot;knowledge through AI”</h3><p>If you&#39;re a researcher, policymaker, or building AI products, here&#39;s a <strong>five-factor warrant profile</strong> that synthesises the literature:</p><ol><li><strong>Method-level reliability (externalist core).</strong> Demonstrate out-of-sample performance, calibration, and error controls appropriate to the domain; monitor <strong>statistically insignificant yet serious</strong> (SIS) failures.</li><li><strong>Provenance &amp; traceability.</strong> Track data lineage, updates, and evaluation cards to ensure testimonial uptake isn&#39;t resting on poisoned or circular corpora. (Supports artificial testimony standards.)</li><li><strong>Opacity management.</strong> If internals aren&#39;t surveyable, replace &quot;interpretability&quot; demands with <strong>process reliability</strong> (validation regimes, adversarial stress-tests, uncertainty exposure).</li><li><strong>Mechanistic affordance.</strong> Where feasible, use mechanistic interpretability or surrogate explanations that link structure to domain mechanisms; otherwise, classify outputs as instrumentally reliable but <strong>non-understanding</strong>.</li><li><strong>Social embedding &amp; accountability.</strong> Treat AI as an epistemic instrument governed by institutions — peer challenge, red-team testimony, escalation paths — rather than as an autonomous expert. This aligns with critiques of conferring <strong>epistemic authority</strong> on AI.</li></ol><h3>7) Answering the title question directly</h3><p>So, <em>what is the nature of knowledge produced by AI</em>?</p><ul><li><strong>Strictly speaking, AI does not (yet) produce &quot;knowledge&quot; as a subject that believes and is answerable for truth.</strong> But it does <strong>produce knowledge for us</strong> when (i) its outputs are the reliable products of well-governed computational methods (instrumental route), or (ii) its utterances are admissible as <strong>artificial testimony</strong> under engineered and institutional reliability standards, or (iii) it contributes to <strong>understanding</strong> by revealing mechanisms or principled patterns. Those are established, defensible channels in mainstream epistemology and the philosophy of science.</li></ul><p>AI doesn&#39;t need to <em>be</em> a knower to <em>produce</em> knowledge. It needs to be the right kind of <strong>instrument</strong> in the correct type of <strong>institution</strong> — reliable in method, disciplined in testimony, and, where possible, connected to mechanisms. That&#39;s how we integrate artificial systems into a <em>human</em> epistemic enterprise.</p><h3>References</h3><p>Durán, J. M., &amp; Formanek, N. (2018). Grounds for trust: Essential epistemic opacity and computational reliabilism. <em>Minds and Machines, 28</em>(4), 645–666. <a href="https://doi.org/10.1007/s11023-018-9481-6">https://doi.org/10.1007/s11023-018-9481-6</a></p><p>Ferrario, A., Panigutti, C., Ratti, C., &amp; Ciuccarelli, P. (2024). Experts or authorities? A conceptual investigation of the epistemic role of AI systems. <em>Philosophy &amp; Technology, 37</em>(2), 1–25. <a href="https://doi.org/10.1007/s13347-024-00729-8">https://doi.org/10.1007/s13347-024-00729-8</a></p><p>He, J. (2025). Testimony by large language models. <em>AI &amp; Society, 40</em>(1), 113–125. <a href="https://doi.org/10.1007/s00146-025-01739-2">https://doi.org/10.1007/s00146-025-01739-2</a></p><p>Heersmink, R. (2024). A phenomenology and epistemology of large language models. <em>Ethics and Information Technology, 26</em>(1), 1–14. <a href="https://doi.org/10.1007/s10676-024-09798-1">https://doi.org/10.1007/s10676-024-09798-1</a></p><p>Humphreys, P. (2009). The philosophical novelty of computer simulation methods. <em>Synthese, 169</em>(3), 615–626. <a href="https://doi.org/10.1007/s11229-008-9435-2">https://doi.org/10.1007/s11229-008-9435-2</a></p><p>Páez, A. (2024). Epistemology of AI systems: Artificial testimony, deference, and epistemic authority. <em>Philosophy of Science, 91</em>(5), 841–857. <a href="https://doi.org/10.1017/psa.2023.45">https://doi.org/10.1017/psa.2023.45</a></p><p>Ratti, C., &amp; Resch, B. (2024). Mechanistic interpretability and the epistemology of AI explanations. <em>Biology &amp; Philosophy, 39</em>(4), 1–22. <a href="https://doi.org/10.1007/s10539-024-09968-0">https://doi.org/10.1007/s10539-024-09968-0</a></p><p>Vallor, S. (2023). The AI mirror: How to reclaim our humanity in an age of machine thinking. <em>Oxford University Press.</em> <a href="https://doi.org/10.1093/oso/9780197672920.001.0001">https://doi.org/10.1093/oso/9780197672920.001.0001</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b58d6c2784a4" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>