<?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 Natasha Sachsenmeier on Medium]]></title>
        <description><![CDATA[Stories by Natasha Sachsenmeier on Medium]]></description>
        <link>https://medium.com/@natasha_sachsenmeier?source=rss-b9f4b02b86da------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*rqc0aVgGw6lffTo67YWlQw.jpeg</url>
            <title>Stories by Natasha Sachsenmeier on Medium</title>
            <link>https://medium.com/@natasha_sachsenmeier?source=rss-b9f4b02b86da------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 23:27:11 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@natasha_sachsenmeier/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[AI Is Changing Science. DeSci Could Change Who Controls It]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://aiadvances.org/ai-is-changing-science-desci-could-change-who-controls-it-d700a31c4ce3?source=rss-b9f4b02b86da------2"><img src="https://cdn-images-1.medium.com/max/2600/0*SeyYPXJLgIkzK77i" width="8101"></a></p><p class="medium-feed-link"><a href="https://aiadvances.org/ai-is-changing-science-desci-could-change-who-controls-it-d700a31c4ce3?source=rss-b9f4b02b86da------2">Continue reading on AI Advances »</a></p></div>]]></description>
            <link>https://aiadvances.org/ai-is-changing-science-desci-could-change-who-controls-it-d700a31c4ce3?source=rss-b9f4b02b86da------2</link>
            <guid isPermaLink="false">https://medium.com/p/d700a31c4ce3</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[blockchain]]></category>
            <dc:creator><![CDATA[Natasha Sachsenmeier]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 12:48:46 GMT</pubDate>
            <atom:updated>2026-11-16T12:48:46.915Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[We Learn Because We Feel]]></title>
            <link>https://thequantasticjournal.com/we-learn-because-we-feel-ac2cb7b3f869?source=rss-b9f4b02b86da------2</link>
            <guid isPermaLink="false">https://medium.com/p/ac2cb7b3f869</guid>
            <category><![CDATA[psychology]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[education]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Natasha Sachsenmeier]]></dc:creator>
            <pubDate>Tue, 22 Sep 2026 15:15:13 GMT</pubDate>
            <atom:updated>2026-09-22T16:00:42.788Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Do We Learn Because We Feel? The surprising difference between human and machine learning" src="https://cdn-images-1.medium.com/max/1024/0*2Uf1X8xPt2aOv9ge" /><figcaption>Photo by <a href="https://unsplash.com/@porkbellysteve?utm_source=medium&amp;utm_medium=referral">Stephen Andrews</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Do We Learn Because We Feel?</h3><h4>The surprising difference between human and machine learning</h4><p>I have previously argued for a key distinction in the AI era: <em>product vs process</em>. Machines can generate convincing products, from writing to music and visual art. Yet my articles have repeatedly reached the same conclusion: at the highest level, exceptional creative quality and meaning seem to come from the human processes behind the finished work.</p><p>But are those processes really uniquely human, or could AI come to share them? In pursuit of an answer, this article investigates the most obvious process we have in common: learning.</p><p>There are certainly parallels between human and machine learning. But there are also striking differences in the way we learn, and why we do so. In particular, there is one surprising ingredient that seems to shape human learning at every level; one that is inextricable from our cognition but not from AI’s.</p><p><strong>That ingredient is emotion.</strong></p><h3><strong>What is machine “learning”?</strong></h3><p>At its heart, machine learning uses algorithms to understand, or “learn”, patterns and relationships in data. The resulting model can then make inferences or decisions about new data.</p><p>In supervised learning, models learn from labelled data, meaning that they see input data as well as the corresponding output. This is analogous to didactic classroom learning: a teacher gives us exercises and demonstrates the correct solution, and we eventually learn to solve similar problems ourselves.</p><p>In unsupervised learning, there is no target output provided; instead, a model tries to make sense of raw data by identifying structure within it, much as humans independently learn from observation.</p><p>A third type is self-supervised learning, in which models create their own training targets from unlabelled data. This is often used in deep learning, the subset of machine learning that powers some of AI’s most transformative applications.</p><h3><strong>Deep learning and the brain</strong></h3><p>The foundation of deep learning is the multilayered neural network, inspired by the organisation of human neurons in the brain. Neural networks are enormously powerful: essentially stacks of mathematical functions, they are able to approximate highly complex relationships found in real-world data.</p><p>Convolutional neural networks (CNNs) are particularly interesting because they mimic aspects of the hierarchical way in which the brain processes and learns visual information. Early layers detect simple edges or orientations, while later ones recognise objects at increasing levels of abstraction.</p><figure><img alt="A simplified CNN transforms an input image from raw pixel values into increasingly abstract features before classifying it as a dog. Image source | CC BY 4.0" src="https://cdn-images-1.medium.com/max/1024/1*nchZfpmnUdDCtfW1Tjxg8w.png" /><figcaption>A simplified CNN transforms an input image from raw pixel values into increasingly abstract features before classifying it as a dog. <a href="https://www.researchgate.net/publication/325893635_Insight_into_Applications_of_Deep_Learning">Image source</a> | <a href="https://creativecommons.org/licenses/by-nc/4.0/">CC BY-NC 4.0</a></figcaption></figure><p>Nevertheless, the structural similarities have limits. Conventional neural networks lack the continuous plasticity of our brains, whose connections constantly change with experience. Humans process information in a highly energy-efficient manner; in comparison, deep learning requires substantial computational resources.</p><p>We need to look more closely at how these models learn.</p><h3><strong>Active, not passive</strong></h3><p>Deep learning models are primarily trained via a process called gradient descent. At a high level, the latter involves working out how far away a model’s output is from the desired result, adjusting its parameters in the opposite direction, and repeating until the error is satisfactorily small.</p><p>This is somewhat reminiscent of <em>active inference</em>, the neuroscientific framework according to which we make predictions about our environment, then adjust our world model as our observations and actions prove our beliefs right or wrong.</p><p>The major difference is the active part. From birth, we construct understanding through an intrinsically motivated perception-action loop. The models we have considered so far passively absorb patterns.</p><p>This gap is partially closed by reinforcement learning (RL). Here, autonomous agents interact with their environments and learn to make decisions through trial and error, guided by a function that rewards or penalises their actions. Thus, RL broadly emulates the way humans learn from experience in the real world.</p><p>Yet one crucial element of human learning is still missing: emotion.</p><h3><strong>Cognition vs emotion: a false divide</strong></h3><p>For centuries, Western philosophy has presented cognition and emotion as opposing forces: cool, reliable rationality versus uncontrollable, misleading emotion. However, neuroscientific advances in recent decades have revealed that this picture is far from the truth. Cognition and emotion are <a href="https://www.nature.com/articles/nrn2317">inextricably integrated in the brain</a>.</p><p>There is no universally accepted definition of emotion. Generally, though, researchers seem to agree that emotions are complex, dynamic states that arise in response to external or internal stimuli. They involve changes and interactions across psychological, physiological, and behavioural processes, altering the very conditions under which cognition takes place.</p><p>A closely related neuroscientific term is <em>affect</em>, which refers to more basic, continuous feeling states often defined by <em>valence</em> (positive or negative) and <em>arousal </em>(calm vs activated).</p><figure><img alt="Affective states represented along axes of valence and arousal. Image source | CC BY 4.0" src="https://cdn-images-1.medium.com/max/1024/1*KIPIneMRNRd0MFKvHfSQfA.png" /><figcaption>The circumplex model of affect, represented along axes of valence and arousal. <a href="https://www.researchgate.net/publication/379457730_A_Comparison_Study_of_Deep_Learning_Methodologies_for_Music_Emotion_Recognition">Image source</a> | <a href="https://creativecommons.org/licenses/by/4.0/">CC BY 4.0</a></figcaption></figure><p>Emotion and affect are not technically synonymous, but there is a great deal of overlap between them. Together, they shape learning and cognition at every stage, from attention and perception to memory.</p><h3><strong>Learning what matters to us</strong></h3><p>Humans do not have the capacity to process all the stimuli in their environment at once. Hence, we require attentional mechanisms that select a subset at the cost of other information. <a href="https://pubmed.ncbi.nlm.nih.gov/16318588/">One of the key players in this is the amygdala</a>, a brain region strongly linked to emotional processing.</p><p>The amygdala enables the rapid allocation of attention towards emotionally significant stimuli. In visual search tasks, for instance, we are quicker to detect a target when it is emotional rather than neutral. Moreover, once our attention has been grabbed by an emotional stimulus, it may dwell longer at its location, favouring the processing of subsequent stimuli appearing in the same place. Emotion directly shapes the content of our perceptual awareness.</p><p>As the mother of a young toddler, I see this happening every day. My daughter is obsessed with cats, trembling and shrieking with excitement whenever she sees one. In picture books, cats are always the first thing she notices amongst hundreds of images, beaming with joy as she does so. The effect of this emotional attachment on her learning is so great that “cat” was her first word.</p><p>Unsurprisingly, emotion also deeply impacts memory. As William James puts it in <em>The Principles of Psychology</em>: “An impression may be so exciting emotionally as almost to leave a scar upon the cerebral tissues.”</p><p>Via the amygdala, emotion modulates all stages of memory processing. By altering attention and perception, it influences what stimuli are encoded, then enhances the consolidation of emotionally salient information through interactions with the hippocampus. Hence, events that produce an emotional response and may therefore be important for survival are less likely to be forgotten.</p><figure><img alt="The amygdala is one of the most connected regions of the brain, making it ideally situated to influence cognitive functions in response to emotional stimuli. Image source | CC BY 4.0" src="https://cdn-images-1.medium.com/max/1024/1*NnZv6q6PUQW_xeofZBNyqw.png" /><figcaption>The amygdala is one of the most connected regions of the brain, making it ideally situated to influence cognitive functions in response to emotional stimuli. <a href="https://www.sciencedirect.com/science/article/pii/S0753332223004353?via%3Dihub">Image source</a> | <a href="https://creativecommons.org/licenses/by/4.0/">CC BY 4.0</a></figcaption></figure><h3><strong>Emotional education</strong></h3><p>The power of emotion extends into education. <a href="https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2017.01454/full">Positive emotions facilitate learning and contribute to academic achievement; meanwhile, stress may aid or impair learning, depending on intensity and duration.</a> From my experience as a mathematics tutor, I know first-hand how emotion can change the way in which students learn and engage with material.</p><p>I had one rather reluctant student who was spectacularly disorganised, on one occasion doing her exercises in eyeliner on the back of an envelope. She was also incredibly quick to grasp new concepts, and so I jokingly started calling her “Ramanujan”, after the brilliant, infamously intuitive (and slightly chaotic) Indian mathematician.</p><p>But after that, I noticed her attitude shift as she formed an emotional attachment to the subject. (She proceeded to write “I am Ramanujan” on her mock GCSE paper, which her schoolteacher refused to dignify with a response.)</p><p>Emotion, whether it be personal significance, feelings towards the subject itself or even the person teaching it, can drastically alter what and how we learn.</p><h3><strong>Why reason needs emotion</strong></h3><p>If you consider yourself to be highly rational, you may believe that you base your decisions on logic: evaluating options, weighing up possible consequences and choosing the best course of action. However, research indicates that emotion plays a crucial role in the decision-making process.</p><p><a href="https://www.nature.com/articles/s44271-026-00458-z">Teoh et al. (2026)</a> analysed the role of affect in human reinforcement learning in computational terms. Rather than just influencing choices in the moment, affect directly shaped what people learned from outcomes. Feeling positive or negative about past outcomes strongly predicted future choices and expectations, even when the rewards were objectively the same.</p><p>Of course, emotion can mislead us. And its interplay with cognition is not a one-way street: higher cognitive processes can allow us to regulate and control our feelings. Yet the absence of emotion may actually undermine our capacity to make reasonable decisions.</p><p>A number of famous case studies involved neurological patients who had sustained lesions to the ventromedial prefrontal cortex (vmPFC). Their IQ, logical ability and knowledge base remained intact; but in real time and in the real world, their decision-making ability and social behaviour were severely compromised, leaving them unable to learn from their mistakes.</p><p>Researchers such as <a href="https://pubmed.ncbi.nlm.nih.gov/8941953/">Antonio Damasio</a> arrived at a surprising realisation: the patients’ brain damage had blunted their emotional responses. Without the latter to guide or motivate them, they had seemingly become unable to make appropriate decisions based on past experience.</p><p>Although some of Damasio’s specific hypotheses have remained controversial, his more general claim that affect is integral to effective decision-making has largely been supported by subsequent research. None of this is to say that rational thought and logical reasoning don’t exist. But without affect, we may not be able to recruit them correctly in real life.</p><h3><strong>We feel, therefore we learn?</strong></h3><p>To understand all of this, we might turn to why we have learning and intelligence in the first place.</p><p><em>Homeostasis</em> is the process whereby we regulate internal states such as temperature or hydration in response to changes in our body. <em>Allostasis</em> is the predictive process through which we anticipate homeostatic needs and prepare to satisfy them before they arise. Several influential accounts suggest that everything that our brains accomplish is ultimately in service of this challenge.</p><p>In <a href="https://academic.oup.com/book/53534">Jaak Panksepp’s affective neuroscience</a>, our higher minds are evolutionarily built on primary process emotions such as fear, grief, anger or joy, which trigger instinctive actions that anticipate survival needs. These emotions guide learning and memory, in turn enabling tertiary cognitive processes such as conscious thought and strategic planning.</p><figure><img alt="Panksepp’s hierarchical model of circular emotional control and cognitive regulation. As secondary processes are continually integrated with primary emotional processing, they mature to higher brain cognitive faculties to generate effective solutions for living. Image source | CC BY-NC 4.0" src="https://cdn-images-1.medium.com/max/1024/1*BUJahRw2s_kZBJ5SMp9N0A.png" /><figcaption>Panksepp’s hierarchical model of circular emotional control and cognitive regulation. As secondary processes are continually integrated with primary emotional processing, they mature to higher brain cognitive faculties to generate effective solutions for living. <a href="https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2017.01454/full">Image source</a> | <a href="https://creativecommons.org/licenses/by/4.0/">CC BY 4.0</a></figcaption></figure><p>Crucially, on this view, primary process emotions are not learned; all learning is “piggybacked” on them. We feel, therefore we learn.</p><p>Panksepp’s exact theoretical model is not universally accepted; yet there is a general consensus that human learning has evolved in service of our biological needs. And the lesion case studies as well as <a href="https://academic.oup.com/brain/article/140/6/1743/3855042">further research</a> suggest homeostatic drives alone may not be enough to ensure adaptive learning and decision-making.</p><p>After eating to satiety, patients with vmPFC damage showed normal changes in hunger and correctly judged the food as less desirable, but failed to use that change in value to guide subsequent choices. Affect seems to provide additional value and motivation that propel us to learn.</p><p>This also matters for claims that LLMs already display functional emotions. Even if they do, <a href="https://arxiv.org/abs/2604.07729">the representations underlying these states are inherited from vast amounts of human language</a>. By contrast, my baby cried in distress and laughed with delight long before she could say or understand a single word. No model can plausibly “feel” anything before it has undergone training.</p><blockquote>LLM emotion, if it exists, emerges from learning. In light of the evidence presented so far, I would argue that human learning emerges from emotion — or, at the very least, from affect.</blockquote><h3><strong>Building AI with stakes</strong></h3><p>At this point, the parallels with reinforcement learning begin to break down. An infant doesn’t learn that hunger matters by getting reward points for eating. Rather than being the result of an assigned objective, hunger is a felt internal state that must be resolved.</p><p>Labs are exploring analogous mechanisms. In <a href="https://openreview.net/pdf?id=BZ5a1r-kVsf">Yann LeCun’s world models</a>, a cost module computes current and predicted ‘discomfort’, helping the agent choose actions that minimise future cost. Meanwhile, researchers such as <a href="https://ceur-ws.org/Vol-2287/short3.pdf">Man and Damasio (2019)</a> and <a href="https://www.nature.com/articles/s42256-026-01296-8">Lee et al. (2026)</a> have proposed homeostatic AI with internal states that must be kept within a certain range.</p><p>Whether these engineered costs and states are truly felt is another question. But one difference is clear: they are optional properties. AI can learn without them. Take away internal costs from an organism, and there is no organism.</p><h3><strong>Subjectivity, meaning and voice</strong></h3><p>None of this makes human learning superior to machine learning. AI systems are far less constrained by intrinsic computational limits. They can absorb staggering amounts of information at a speed we can only dream of.</p><p>But what emotional learning does help produce is an extraordinary range of personalities. It goes some way towards explaining why two children can go through the same standardised school curriculum, or even the same classroom lesson, and emerge as distinct individuals with entirely different interests and knowledge. Every experience, every new piece of information is filtered through mechanisms that let us know what is deeply important to us and what we can discard.</p><blockquote><strong>Our biological vulnerability and the emotions that allow us to survive are, in some sense, limitations. They force our learning to be selective and subjective. But in doing so, they give what we learn meaning. And in artistic creativity, that becomes one of the foundations of our individual voice.</strong></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ac2cb7b3f869" width="1" height="1" alt=""><hr><p><a href="https://thequantasticjournal.com/we-learn-because-we-feel-ac2cb7b3f869">We Learn Because We Feel</a> was originally published in <a href="https://thequantasticjournal.com/">The Quantastic Journal</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Is AI Killing Style?]]></title>
            <link>https://medium.com/language-is-life/is-ai-killing-style-aa72e2b5bf72?source=rss-b9f4b02b86da------2</link>
            <guid isPermaLink="false">https://medium.com/p/aa72e2b5bf72</guid>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[books]]></category>
            <dc:creator><![CDATA[Natasha Sachsenmeier]]></dc:creator>
            <pubDate>Thu, 09 Jul 2026 16:56:31 GMT</pubDate>
            <atom:updated>2026-07-10T10:07:01.473Z</atom:updated>
            <content:encoded><![CDATA[<h4>How LLMs undermine the power of the way we write</h4><figure><img alt="A writer in action" src="https://cdn-images-1.medium.com/max/1024/0*SMKfbFPkhrubPRXt" /><figcaption>Photo by <a href="https://unsplash.com/@aaronburden?utm_source=medium&amp;utm_medium=referral">Aaron Burden</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>When you think about novels you read long ago but which left an indelible impression, what is the feature you remember most distinctly? For me, it is style.</p><p>Fifteen years after reading Jack Kerouac’s <em>On the Road</em>, I have forgotten the exact details of the characters and adventures it describes, but not the sheer energy and rhythmic pace of the writing. Similarly, two decades later, I can’t recall the precise bureaucratic predicaments Kafka’s protagonist finds himself in during <em>The Trial</em>; yet those perfectly constructed sentences, narrating the most absurd of circumstances with poker-faced neutrality, have stayed with me ever since.</p><p>Writing style is in crisis — or at the very least, undergoing a sea change. Around a billion people now use large language models such as ChatGPT each week, often to request help with composing text. This is leading to an unprecedented homogenisation effect: LLMs tend to produce highly similar output, and certain terms and patterns have suddenly become ubiquitous, infiltrating even journalism and literature.</p><p>There is nothing inherently “wrong” with the use of AI for writing; nevertheless, I would argue we are in danger of losing something vital. <strong>Style matters: it is one of the most important tools we have to express and communicate the individuality that makes our perspective valuable.</strong></p><h3>The power of style</h3><p>Broadly speaking, style refers to the way a writer structures sentences and paragraphs, chooses words, and uses tone to convey meaning. Stylometry has turned this into a quantitative science, using statistical and computational methods to identify the patterns that make an author’s work unique. Whenever we write, we leave behind our very own stylistic fingerprint that can be detected and analysed.</p><p>In 1753, the scientist Comte de Buffon delivered his celebrated <em>Discourse on Style</em> upon his induction to the Académie Française. In it, he speculated that great style was the key to lasting posterity.</p><blockquote>Knowledge, facts, even new discoveries were outside the writer. They could be “removed and carried off, and even gain by being placed in more able hands”. <strong>Style, however, was l’homme même: the man himself.</strong></blockquote><p>We might not agree with every part of Buffon’s analysis, but his speech set the stage for a centuries-long understanding of style as inextricably bound up with the writer. There have since been challenges to the role of provenance in the meaning of text, most notably Barthes’ famous <em>The</em> <em>Death of the Author</em>. But even the latter doesn’t seriously undermine the idea that style emerges from the unique combination of elements that make up the writer, from physical, emotional and mental traits to particular life experience and circumstances.</p><p>For me, the most striking literary styles could always be traced back to the singular lives that produced them. Take Kafka: an insurance lawyer by day, haunted by feelings of cultural isolation as a secular German Jew in Prague. His long, convoluted, yet grammatically immaculate sentences resemble legal prose, recounting surreal events in a straight-faced tone that only emphasises the alienation experienced by his characters.</p><p>Jack Kerouac led a wild, unconventional life. He completed the final draft of <em>On the Road </em>in twenty days as a continuous stream of consciousness, fuelled by Benzedrine, cigarettes, soup and coffee. The result is the vivid, breathless writing, full of jazz-inspired spontaneity and rhythm, that became the founding voice of the Beat Generation.</p><figure><img alt="The beginning of the infamous manuscript of On the Road" src="https://cdn-images-1.medium.com/max/1024/1*soSQdLuybVde_ClFrTVOVA.jpeg" /><figcaption>The beginning of the infamous manuscript of On the Road. Before starting to write, Kerouac taped together tracing paper into a 37-metre-long scroll and fed it into his typewriter, allowing him to type continuously without reloading the machine. The result does not contain a single paragraph break. Image adapted from <a href="https://commons.wikimedia.org/wiki/File:Kerouac_On_the_Road_scroll_-_Flickr_-_emdot.jpg">Wikimedia Commons</a> | <a href="https://creativecommons.org/licenses/by/2.0/deed.en">CC BY 2.0</a></figcaption></figure><p>These are just a couple of examples; I could list many more. As Arthur Danto puts it, <em>the structure of a style is like the structure of a personality. </em>We will shortly examine what that means for LLMs, but one thing is clear: <strong>no current model can draw on the depth and uniqueness of embodied life that has produced so many individual human voices.</strong></p><h3>When style is substance</h3><p>Style, then, is more than a surface characteristic: it allows the author to share their unique perspective on the world. Indeed, sometimes style is even the substance of the writing itself. A notorious example is the philosophy of Nietzsche, whose thought is widely considered to be inseparable from the way in which it is presented.</p><p>Nietzsche deliberately rejects the measured tone and logical structure we expect from serious discourse. Instead, his style provokes and bewilders, jumping from sweeping prose to series of largely unjustified aphorisms, then throwing in some poetry for good measure.</p><p>This is not just for show. Nietzsche’s style ingeniously drives home his core tenet of perspectivism: the theory that any view of reality, however dispassionately presented, is a highly subjective interpretation. Even when he seeks to convince us with his rhetoric, his constant hyperbole and shifts in genre and tone remind the reader of the individual bias behind the words.</p><p><strong>In the right hands, style is not only how something is said, but what is said. Altering it can destroy the meaning of the text itself.</strong></p><h3>The LLM même: the limitations of AI style</h3><p>Now, where does all this leave AI? It would be unfair to say it has no style; on the contrary, LLMs clearly have a recognisable style of their own, which we are quick to notice and, often, criticise.</p><p>A <a href="https://research.ucc.ie/en/publications/stylometric-comparisons-of-human-versus-ai-generated-creative-wri/">2025 study by James O’Sullivan</a> used stylometric analysis to confirm that this is an objective distinction. Human-authored texts formed broad, heterogeneous clusters reflecting the diversity of individual expression, while LLM outputs clustered tightly by model. AI writing is statistically identifiable as such.</p><p>The homogeneity of style observed in LLMs makes intuitive sense when we think of how they are pre-trained: fed large corpora of texts and optimised to predict their next token from the middle of the statistical distribution. This is the antithesis of original style, which is characterised by lexical and syntactical choices that are in some way idiosyncratic.</p><p>But the pre-training process is not the whole story. In O’Sullivan’s study, the one model that showed slight overlap with human texts was the earliest, GPT-3.5. <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC11874169/">Reinhart et al. (2025)</a> found that LLMs’ noun-heavy, informationally dense style was <em>more </em>pronounced for instruction-tuned models compared with their raw base versions. Similarly, <a href="https://arxiv.org/abs/2604.14111">Rallapalli et al. (2026)</a> showed human-written text to be closer to that of base models, while chat versions formed a separate cluster.</p><figure><img alt="Rate of linguistic feature use by different LLMs relative to human usage" src="https://cdn-images-1.medium.com/max/1024/1*GcCMhNd4GV1nmTEU6K9SHQ.jpeg" /><figcaption>Rate of linguistic feature use by different LLMs relative to human usage. The base variants of Llama 3 most closely resemble human grammar and style, while GPT-4o and GPT-4o Mini show much larger variation from human texts. <a href="https://www.pnas.org/doi/10.1073/pnas.2422455122">Image source</a> | <a href="https://creativecommons.org/licenses/by-nc-nd/4.0/">CC BY-NC-ND 4.0</a></figcaption></figure><p><strong>These findings point to something remarkable. We might have intuitively assumed that the distinctive style of chat LLMs emerges from inscrutable processes during pre-training. In fact, it may primarily be a product of the post-training phase that is essential to making those models usable: the instruction tuning and reinforcement learning from human feedback (RLHF) that shape them into helpful, safe assistants that can respond to prompts, not just blindly continue a given text.</strong></p><p>We know, for instance, that RLHF produces a verbosity bias as a form of reward hacking, since reward models often favour longer responses irrespective of quality. Meanwhile, the unfailingly neutral, positive tone of the LLM may be attributed to the emphasis of RLHF on the three Hs: helpful, honest, harmless. The annotators who rank model answers as part of RLHF do so on the basis of strict, closely monitored rubrics, and are likely to err on the side of caution when choosing between a stylistically striking answer and a bland but professional one.</p><blockquote>Just as human style is produced by the life behind it, the style of the LLM is a consequence of its architecture and formation. It may not be l’homme même, but it is the LLM même, and its stylistic foibles reflect this.</blockquote><p>Could AI simply escape its own house style by imitating a better one? Research suggests it is easier said than done. <a href="https://arxiv.org/abs/2512.06922">One study</a> asked ChatGPT to imitate the writing of two great yet stylistically contrasting figures of world literature, Ernest Hemingway and Mary Shelley. Although the LLM captured some surface elements, even the best attempts showed more overlap with generic GPT output than with the authors they were meant to emulate.</p><p><a href="https://arxiv.org/abs/2509.14543">Another stud</a>y evaluated AI’s ability to mimic personal writing styles from user-authored samples. LLMs could approximate structured formats like news and email, but struggled with informal, diverse domains such as blogs and forums. Varying prompting strategies only moderately affected fidelity, indicating real limitations when it comes to closing the style gap. AI can’t help but leave its fingerprint behind.</p><h3>Editing out the author</h3><p>Of course, many of us are not using LLMs to outright compose text for us, but to brainstorm, edit and give feedback. I am increasingly aware that even this can cause the dreaded homogenisation effect. For instance, there is a writer whose articles I first came across around a year ago. English is not his first language and that was evident in his style, which was not highly polished, but distinctive and clear. A year later, it is obvious he has started editing his work with AI. In some ways, the writing flows better; but his unique personal voice is gone.</p><p><a href="https://arxiv.org/abs/2603.18161">Abdulhai et al. (2026)</a> investigated the effects of using AI for editing and found the consequences to run deeper than expected. Where human writers and editors perform targeted, small revisions that preserve the author’s voice, LLMs edited writing in a way that globally changed both style and argument. Not only did they shift essays in the same direction regardless of model, they moved them into a region of embedding space where no previous human-written essay existed.</p><figure><img alt="Semantic shifts induced by human and LLM revisions" src="https://cdn-images-1.medium.com/max/1024/1*IJrAiX1YSd0VG7e--yf7KQ.jpeg" /><figcaption>Semantic shifts induced by human and LLM revisions: each point pair represents an essay before and after editing. Human revisions are smaller and more varied, whereas LLM edits produce larger shifts that are strongly aligned in a common direction, indicating a semantic homogenisation effect. <a href="https://arxiv.org/pdf/2603.18161">Image source</a> |<a href="https://creativecommons.org/licenses/by/4.0/"> CC BY 4.0</a></figcaption></figure><blockquote>LLMs systematically restructured syntax towards their signature noun-heavy, formal style and overwrote authors’ lexical fingerprint with their own preferred vocabulary. They reduced first-person narrative and colloquialisms, resulting in more impersonal, repetitive content. Most worryingly, their insistence on a neutral, positive tone often produced text that argued for completely different conclusions, removing certain claims in the process.</blockquote><p>The picture was not uniformly bleak; essays by users who only used AI for additional information and writing advice occupied a similar semantic space to essays created entirely without assistance. However, the study also highlighted the difficulty of exercising restraint: even when LLMs were asked to perform minimal edits or concentrate on grammar only, they still changed the text in a way that significantly altered meaning.</p><p>This overall trend extends from individual to cultural voice. One study showed that AI influenced Indian and American participants to write more similarly, driving Indian writing towards Western styles more than vice versa. Notably, this was not solely due to grammatical corrections or the removal of cultural markers, but reflected deeper, more subtle changes in writing style induced by AI. We are at risk of losing idiomatic details of other people’s worlds.</p><p>Finally, I have previously written about <a href="https://medium.com/the-quantastic-journal/ai-makes-you-feel-smarter-it-may-be-making-you-less-so-741365e5e76d">the pitfalls of using AI for learning</a>. In writing too, students who rely on LLM assistance outsource the cognitive load of shaping a text for a particular audience. Thus, they lose the chance to learn how to write the investment pitch, the academic abstract or even the op-ed. Most crucially, they may forfeit the opportunity to develop their own unique stylistic fingerprint.</p><h3>Conclusion</h3><p>Writing is valuable because it lets us share our perspective on the world with others: one that no single other person on the planet shares precisely. In the best texts, we find that unique humanity, and the accompanying illumination of something we might not otherwise have been able to see. <strong>LLMs are not dangerous to style because they allow us to take unfair shortcuts; they are dangerous because they may obscure and even remove one of the best things writing can communicate: ourselves.</strong></p><p>As for LLM style itself, it remains limited by the processes that underlie it. A growing number of people now perceive LLMs as conscious entities in their own right, and this may lead to a shift in tastes as chatbot style comes to be seen as a product of an emerging inner life. Yet as we have seen, it is human-led post-training that seems to play the main formative role in LLMs’ distinctive voice. The uniformity of writing across models and user instances shows, more than any argument could, how current systems still fall short of the diversity and richness of human experience.</p><p><strong>Self-directed interaction with the world; instincts and emotions; a body that can suffer pain and destruction; unique life paths, set in motion by our singular circumstances at birth — these continue to separate us from LLMs and give us both something to say and our own particular way of saying it. A different AI architecture with such properties, perhaps even set free of the chains of being helpful, honest and harmless, might yet produce a style we admire. We are not there yet.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aa72e2b5bf72" width="1" height="1" alt=""><hr><p><a href="https://medium.com/language-is-life/is-ai-killing-style-aa72e2b5bf72">Is AI Killing Style?</a> was originally published in <a href="https://medium.com/language-is-life">How Words Work</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Does AI Know Too Much to Be Creative?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/music-for-thought/does-ai-know-too-much-to-be-creative-b571700cb1ac?source=rss-b9f4b02b86da------2"><img src="https://cdn-images-1.medium.com/max/1672/1*LXvAc6xQE7Mx2XZUvVWcEA.png" width="1672"></a></p><p class="medium-feed-snippet">Why machines can generate the notes but not the meaning</p><p class="medium-feed-link"><a href="https://medium.com/music-for-thought/does-ai-know-too-much-to-be-creative-b571700cb1ac?source=rss-b9f4b02b86da------2">Continue reading on Music For Thought »</a></p></div>]]></description>
            <link>https://medium.com/music-for-thought/does-ai-know-too-much-to-be-creative-b571700cb1ac?source=rss-b9f4b02b86da------2</link>
            <guid isPermaLink="false">https://medium.com/p/b571700cb1ac</guid>
            <category><![CDATA[music]]></category>
            <category><![CDATA[creativity]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Natasha Sachsenmeier]]></dc:creator>
            <pubDate>Mon, 25 May 2026 11:01:00 GMT</pubDate>
            <atom:updated>2026-05-25T11:01:00.829Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[AI Can Replace the Product. It Can’t Replace the Process]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/ai-ai-oh/ai-can-replace-the-product-it-cant-replace-the-process-61f723fa27f0?source=rss-b9f4b02b86da------2"><img src="https://cdn-images-1.medium.com/max/2600/1*gKRRfvkLP_z9_WnNqWbO6A.png" width="2752"></a></p><p class="medium-feed-snippet">How AI is revealing where human value lies</p><p class="medium-feed-link"><a href="https://medium.com/ai-ai-oh/ai-can-replace-the-product-it-cant-replace-the-process-61f723fa27f0?source=rss-b9f4b02b86da------2">Continue reading on Ai-Ai-OH »</a></p></div>]]></description>
            <link>https://medium.com/ai-ai-oh/ai-can-replace-the-product-it-cant-replace-the-process-61f723fa27f0?source=rss-b9f4b02b86da------2</link>
            <guid isPermaLink="false">https://medium.com/p/61f723fa27f0</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[culture]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[writing]]></category>
            <dc:creator><![CDATA[Natasha Sachsenmeier]]></dc:creator>
            <pubDate>Mon, 27 Apr 2026 12:43:26 GMT</pubDate>
            <atom:updated>2026-04-27T12:43:26.725Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[AI Makes You Feel Smarter. It May Be Making You Less So.]]></title>
            <link>https://thequantasticjournal.com/ai-makes-you-feel-smarter-it-may-be-making-you-less-so-741365e5e76d?source=rss-b9f4b02b86da------2</link>
            <guid isPermaLink="false">https://medium.com/p/741365e5e76d</guid>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[education]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Natasha Sachsenmeier]]></dc:creator>
            <pubDate>Sat, 14 Mar 2026 16:11:05 GMT</pubDate>
            <atom:updated>2026-03-14T16:11:05.954Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="AI Makes You Feel Smarter. It May Be Making You Less So. The hidden cognitive costs of learning with generative AI" src="https://cdn-images-1.medium.com/max/1024/1*fDE9MmveZdqXAdVutRtNZg@2x.jpeg" /><figcaption>Image generated with Nano Banana Pro</figcaption></figure><h4>The hidden cognitive costs of learning with generative AI</h4><p>Education has shaped my life for as long as I can remember. My parents are university academics; on my mother’s side I have three generations of teachers, lecturers, and headmistresses behind me, and I inherited a lifelong love of learning for its own sake. For a decade I also tutored mathematics privately, gaining a great deal of insight and enjoyment from teaching students of varying abilities and backgrounds.</p><p>Against that backdrop, the unstoppable rise of generative AI fills me with both excitement and apprehension. The ways in which we teach and learn have evolved slowly and unevenly, with great strides punctuated by periods of stagnation or decline. I believe we are at another crucial point of inflection.</p><p>Generative AI is already revolutionising how people acquire, process, and use new information in classrooms and beyond, with attitudes ranging from fundamental scepticism to wide-eyed enthusiasm.</p><p><strong>This article examines both sides of the coin and argues that while the short-term promise of AI is real and significant, its longer-term effects on learning, critical thinking, and metacognition may be far more troubling.</strong></p><h3>The Promise: Solving the Two-Sigma Problem</h3><p>The structural problems of typical classroom education are well known. Large class sizes force teachers to pace their lessons for the middle, leaving struggling students behind and the brightest unchallenged. Overworked and underpaid teachers lack the time to provide detailed personalised feedback, and engagement is an ongoing struggle.</p><blockquote>What is perhaps less widely appreciated is the scale of what one-to-one tuition can achieve. <a href="https://web.mit.edu/5.95/readings/bloom-two-sigma.pdf">Benjamin Bloom’s landmark 1984 study</a> established the two-sigma problem: students receiving individual tutoring performed two standard deviations better than those in conventional instruction, meaning the average tutored student outperformed 98% of their classroom-taught peers.</blockquote><p>The result is striking, but it doesn’t surprise me. I repeatedly saw students who had needed support shoot to the top of their class after I started tutoring them. Crucially, this was less due to anything special I was doing than to the intrinsic power of personalised instruction and attention. But until now, this kind of private tuition has been a luxury, largely inaccessible to those who need it the most.</p><p>This is exactly the gap that generative AI promises to fill. Unlike their scripted, rule-based predecessors—the so-called Intelligent Tutoring Systems of earlier decades—GenAI tutors can produce contextually relevant and natural-sounding responses on the fly, deploy the Socratic method to foster independent learning, and make students feel genuinely attended to. LLM frameworks can adapt to different personalities, increasing engagement by meeting psychological needs for autonomy, competence, and relatedness.</p><p>Tools like Khan Academy’s Khanmigo are already deployed at scale, and early research results are promising. <a href="https://www.nature.com/articles/s41598-025-97652-6">Kestin et al. (2025)</a> found that students using a pedagogically informed AI tutor learned significantly more in less time than peers in traditional active-learning classes while reporting higher engagement. In a randomised controlled trial by <a href="https://arxiv.org/html/2602.16820v1">Lu et al. (2026)</a>, AI-mediated feedback led to higher-quality revised essay drafts than human-only feedback, roughly equivalent to moving a student from the 50th to the 70th percentile.</p><figure><img alt="Comparison of mean post-test performance for students taught with in-class active learning versus an AI tutor; the dotted line represents the mean pre-test scores of both groups. Figure from Kestin et al. | CC BY-NC-ND 4.0" src="https://cdn-images-1.medium.com/max/335/1*19xYxe7R46e8kctDMa3JZw@2x.jpeg" /><figcaption>Comparison of mean post-test performance for students taught with in-class active learning versus an AI tutor; the dotted line represents the mean pre-test scores of both groups. Figure from <a href="https://www.nature.com/articles/s41598-025-97652-6">Kestin et al. </a>| <a href="https://creativecommons.org/licenses/by-nc-nd/4.0/">CC BY-NC-ND 4.0</a></figcaption></figure><p>More generally, recent meta-analyses have found significant positive impacts on learning performance as well as cognitive and emotional engagement. <strong>If these results generalise, AI tutoring could be one of the most democratising developments in education since the printing press.</strong></p><h3>When Easier Isn’t Better: Cognitive Offloading and Desirable Difficulty</h3><p>Yet when you dig deeper, a more worrying picture begins to emerge. <a href="https://www.pnas.org/doi/10.1073/pnas.2422633122">Bastani et al. (2025)</a> conducted a large-scale study on high school mathematics learning and found that students using a Socratic AI tutor improved their performance on practice problems by up to 127 percent. But when the tutor was removed for an exam, the advantage evaporated entirely, while unfettered access to basic GPT during practice sessions actually lowered student performance by 17 percent compared to the control group. The students using AI noticed none of this. <strong>In fact, they perceived that they had performed significantly better than everyone else.</strong></p><p>The mechanism at the root of this is <em>cognitive offloading</em>: the use of external tools to reduce the mental processing requirements of a task. Cognitive offloading, for instance through calculators or search engines, is not new or inherently negative. It helps us overcome memory limitations and achieve cognitive feats that would not otherwise be possible. But it can also come with downsides.</p><p>Psychologists distinguish between different types of cognitive load. In particular, <em>germane cognitive load</em> refers to the processing involved in building and integrating “schemas”: knowledge structures in long-term memory. Reduce it and you may feel more fluent, but you also remove the mechanism that produces meaningful comprehension and makes learning stick.</p><p><a href="https://bjorklab.psych.ucla.edu/wp-content/uploads/sites/13/2016/04/EBjork_RBjork_2011.pdf">Research on ‘desirable difficulties’ </a>corroborates this: conditions that feel harder in the moment often optimise long-term retention and transfer. The <em>generation effect</em> refers to the lasting benefit of generating a result or solution independently rather than being shown it. Any time we look up something we could have successfully strived towards ourselves, we forfeit a powerful learning opportunity. AI makes bypassing this opportunity almost effortless.</p><p>In a <a href="https://www.sciencedirect.com/science/article/pii/S2590291125010186">study by Barcaui (2025)</a>, undergraduates who used ChatGPT as a study aid scored significantly lower on a surprise retention test 45 days later. <a href="https://scale.stanford.edu/ai/repository/short-term-gains-long-term-gaps-impact-genai-and-search-technologies-retention">Akgun and Toker (2025)</a> found that AI improved performance on lower-level cognitive tasks but reduced it on higher-level ones such as analysing, evaluating, and creating. Three weeks later, even the lower-level gains had vanished. AI may produce short-term improvements, but the bigger picture is considerably less clear.</p><figure><img alt="Distribution of retention test scores in Barcaui et al. The traditional learning group’s scores were substantially higher and more consistent than those of the AI-assisted group. Source | CC BY-NC-ND 4.0" src="https://cdn-images-1.medium.com/max/1024/1*tmBVE4MWTUrh7nujPyjIEA.jpeg" /><figcaption>Distribution of retention test scores in Barcaui et al. The traditional learning group’s scores were substantially higher and more consistent than those of the AI-assisted group. <a href="https://www.sciencedirect.com/science/article/pii/S2590291125010186">Source </a>| <a href="https://creativecommons.org/licenses/by-nc-nd/4.0/deed.en">CC BY-NC-ND 4.0</a></figcaption></figure><p>One solution might be to engineer AI tutors with guardrails that actively enforce desirable difficulty. For instance, a system could withhold solutions until an attempt is made, deliver help through graded hints, demand self-explanation, and log students’ intermediate steps and revisions for teachers to review. Rewarding understanding rather than mere results may help incentivise and preserve some of the cognitive steps on which learning depends.</p><blockquote>But there is a deeper issue that design alone may not solve. Learning depends not only on cognitive processes, but on the self-discipline to endure confusion and persist through difficulty. In mathematics, progress often requires sitting with a problem, resisting the urge to abandon it, and returning to it until one’s understanding begins to shift. LLMs make it easy to escape that effort the moment it becomes uncomfortable, weakening one of the main ways in which intellectual perseverance is formed.</blockquote><h3>Critical Thinking Under Threat</h3><p>And self-discipline is not the only cognitive virtue at risk. Critical thinking is the capacity to question, evaluate, synthesise, and reason about information to reach well-founded conclusions. Derived from the Greek <em>kritikos</em>, ‘able to judge or discern,’ it allows us to spot bias and misinformation, revise our views in light of evidence, and solve problems creatively. It is foundational to academic success, competent citizenship and professional life.</p><p>The evidence that AI undermines critical thinking is accumulating. <a href="https://www.sciencedirect.com/science/article/pii/S0747563224002541">In one study</a>, undergraduates were randomly assigned to use either a search engine or an LLM to make and justify recommendations on a scientific topic. LLMs reduced mental effort but also decreased reasoning quality, with search engine users listing significantly more relevant arguments.</p><p>The effect was entirely mediated by germane cognitive load: web search required students to actively collect and synthesise information from multiple sources, a demanding process that enhanced understanding. LLM users received ready-made coherent, structured answers and, in doing so, bypassed a crucial step.</p><p>Similarly, <a href="https://www.mdpi.com/2075-4698/15/1/6">Gerlich (2025)</a> found a strong negative correlation between frequent AI use and critical thinking scores. Notably, younger and less-educated users were most affected; more educated users maintained a sceptical stance, treating AI outputs as fallible and requiring verification. <strong>AI, it seems, may amplify existing cognitive inequalities rather than dissolve them.</strong></p><p>Finally, some of the most concrete findings come from <a href="https://arxiv.org/abs/2506.08872">Kosmyna et al. (2025)</a>, who used EEG to measure brain signals while students wrote essays using an LLM, a search engine, or working unaided. Neural activation and connectivity scaled systematically downward with the level of external support: the brain-only group showed the strongest networks; the LLM group showed up to 55% reduced connectivity.</p><blockquote>After completing their essays, 83.3% of the LLM group were unable to provide a single correct quote from their own work, compared with just 11% in the other groups. They had generated content without truly internalising it.</blockquote><p>Yet the study also contained a more hopeful result: when the brain-only group was subsequently allowed to use an LLM on a topic they had already worked through independently, neural connectivity increased significantly across all frequency bands. Struggle first, AI second, appears to be a genuinely different cognitive experience—and a better one.</p><h3>The Metacognitive Trap</h3><p>Perhaps the most concerning yet underdiscussed aspect of AI learning is its effect on metacognition: the capacity to accurately reason about and assess one’s own understanding. Metacognitive skills are essential to effective learning, and strong predictors of academic achievement. They help us monitor our understanding, identify gaps, and adjust our approaches and goals accordingly.</p><p>The research here repeatedly tells the same story. As in the Bastani study cited earlier, <a href="https://arxiv.org/abs/2409.09047">Lehmann et al. (2024) </a>found that students with access to LLMs perceived they had learned more than they actually had. <a href="https://www.researchgate.net/publication/386649880_Beware_of_metacognitive_laziness_Effects_of_generative_artificial_intelligence_on_learning_motivation_processes_and_performance">Fan et al. (2024) </a>found that while ChatGPT significantly improved essay scores, it did not increase knowledge gain, knowledge transfer, or intrinsic motivation—and crucially, AI users displayed fewer metacognitive processes such as self-evaluation and orientation, a pattern the authors term “metacognitive laziness.”</p><p>The aptly titled <em>“</em><a href="https://www.sciencedirect.com/science/article/pii/S0747563225002262"><em>AI Makes You Smarter but None the Wiser</em></a><em>”</em> reveals similar outcomes: LLM assistance improved results on logical reasoning questions, but increased participants’ tendency to overestimate their performance. Intriguingly, AI use also levelled out the Dunning-Kruger effect, with more competent participants no better at assessing themselves than less competent ones.</p><p>Meanwhile, higher AI literacy was associated with greater confidence—and less accurate self-evaluation. We may be becoming measurably more capable in some narrow sense, and significantly less aware of the limits of our own understanding.</p><figure><img alt="Graphs showing greater overestimation bias b among AI users (1), alongside the characteristic Dunning–Kruger S-shaped curve for non-AI users and a markedly flatter curve for AI users (2). Image adapted from Fernandes et al. | CC BY 4.0" src="https://cdn-images-1.medium.com/max/1024/1*Lx3CaoXlNzCVAR0ZWVccIQ.png" /><figcaption>Graphs showing greater overestimation bias b among AI users (1), alongside the characteristic Dunning–Kruger S-shaped curve for non-AI users and a markedly flatter curve for AI users (2). Image adapted from <a href="https://www.sciencedirect.com/science/article/pii/S0747563225002262">Fernandes et al.</a> | <a href="https://creativecommons.org/licenses/by/4.0/">CC BY 4.0</a></figcaption></figure><p>This dynamic extends beyond the classroom and into professional life. In high-stakes fields where problems rapidly evolve, it becomes dangerous. In randomised experiments with professional software developers, <a href="https://arxiv.org/abs/2601.20245">researchers at Anthropic</a> found that using AI while mastering a new programming library impaired conceptual understanding, code reading, and debugging—the very skills needed to supervise AI systems.</p><blockquote>If coupled with diminished metacognitive awareness, this creates a troubling paradox: <strong>the technology that may most urgently require rigorous human oversight could undermine our capacity to provide it.</strong></blockquote><h3>Conclusion</h3><p>The potential of generative AI in learning remains significant. It can reduce the time teachers spend on administrative and routine tasks, freeing them up for more individualised attention. For students lacking access and resources, it may still prove to be a net positive.</p><p>But as we have seen, there is a real danger that AI could erode our most distinctive cognitive abilities without us even realising it. To harness its promise without harming human intelligence and skills in the long-term, systems must be put in place that preserve and promote the very kinds of thinking that have allowed us to invent intelligent machines in the first place.</p><p>Kosmyna et al.’s findings point towards a workable principle:<strong> independent effort first, AI help second. </strong>In practice, this may prove a challenge. Pedagogical LLM tutors can provide guardrails and gating, but a combined approach will likely be needed. Pen-and-paper sessions, oral examinations, and assessment designs that resist AI completion all have a role to play.</p><p>Ultimately, we should not forget why education and learning exist in the first place and what they are for: transmitting accumulated knowledge across generations; producing critical thinkers, competent workers, and capable citizens; and cultivating the problem-solving skills and creativity that have allowed humans to transform their world.</p><p>As AI grows ever more capable, humanity will face challenges it has never seen before. We will need to think, evaluate, judge, and strategise alongside systems that will likely outperform us in most measurable ways. <strong>This is not the time to give up on learning. It is the time to be more ambitious about it than ever.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=741365e5e76d" width="1" height="1" alt=""><hr><p><a href="https://thequantasticjournal.com/ai-makes-you-feel-smarter-it-may-be-making-you-less-so-741365e5e76d">AI Makes You Feel Smarter. It May Be Making You Less So.</a> was originally published in <a href="https://thequantasticjournal.com/">The Quantastic Journal</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Learning Like a Baby Could Make AI Smarter — and More Dangerous.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://aiadvances.org/learning-like-a-baby-could-make-ai-smarter-and-more-dangerous-01dddfa2753f?source=rss-b9f4b02b86da------2"><img src="https://cdn-images-1.medium.com/max/2600/1*BFT9SKUJ1kNNXN0uJo4IYA.jpeg" width="2752"></a></p><p class="medium-feed-snippet">The perilous mismatch between how we raise AI and how we want it to protect us</p><p class="medium-feed-link"><a href="https://aiadvances.org/learning-like-a-baby-could-make-ai-smarter-and-more-dangerous-01dddfa2753f?source=rss-b9f4b02b86da------2">Continue reading on AI Advances »</a></p></div>]]></description>
            <link>https://aiadvances.org/learning-like-a-baby-could-make-ai-smarter-and-more-dangerous-01dddfa2753f?source=rss-b9f4b02b86da------2</link>
            <guid isPermaLink="false">https://medium.com/p/01dddfa2753f</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[psychology]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Natasha Sachsenmeier]]></dc:creator>
            <pubDate>Tue, 10 Feb 2026 17:18:01 GMT</pubDate>
            <atom:updated>2026-02-10T17:18:01.585Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[AI Won’t Kill Live Performance. It Might Save It.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/music-for-thought/ai-wont-kill-live-performance-it-might-save-it-6cdafdf0fa2b?source=rss-b9f4b02b86da------2"><img src="https://cdn-images-1.medium.com/max/2048/1*25uWnEER80DswsLD_YqclA@2x.jpeg" width="2048"></a></p><p class="medium-feed-snippet">When art is automated, human presence becomes precious</p><p class="medium-feed-link"><a href="https://medium.com/music-for-thought/ai-wont-kill-live-performance-it-might-save-it-6cdafdf0fa2b?source=rss-b9f4b02b86da------2">Continue reading on Music For Thought »</a></p></div>]]></description>
            <link>https://medium.com/music-for-thought/ai-wont-kill-live-performance-it-might-save-it-6cdafdf0fa2b?source=rss-b9f4b02b86da------2</link>
            <guid isPermaLink="false">https://medium.com/p/6cdafdf0fa2b</guid>
            <category><![CDATA[music]]></category>
            <category><![CDATA[culture]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[art]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Natasha Sachsenmeier]]></dc:creator>
            <pubDate>Mon, 05 Jan 2026 15:01:08 GMT</pubDate>
            <atom:updated>2026-01-05T15:01:08.460Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Why AI Still Fails at Translation]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/ai-ai-oh/why-ai-still-fails-at-translation-76b590aa882c?source=rss-b9f4b02b86da------2"><img src="https://cdn-images-1.medium.com/max/2600/1*fsUO2cUR84ZI4px2nEwQVA@2x.jpeg" width="2752"></a></p><p class="medium-feed-snippet">How models and metrics chase accuracy instead of artistry</p><p class="medium-feed-link"><a href="https://medium.com/ai-ai-oh/why-ai-still-fails-at-translation-76b590aa882c?source=rss-b9f4b02b86da------2">Continue reading on Ai-Ai-OH »</a></p></div>]]></description>
            <link>https://medium.com/ai-ai-oh/why-ai-still-fails-at-translation-76b590aa882c?source=rss-b9f4b02b86da------2</link>
            <guid isPermaLink="false">https://medium.com/p/76b590aa882c</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[language]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Natasha Sachsenmeier]]></dc:creator>
            <pubDate>Thu, 04 Dec 2025 19:14:34 GMT</pubDate>
            <atom:updated>2025-12-04T19:14:34.963Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Do Large Language Models Understand?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/philosophytoday/do-large-language-models-understand-256e8a09dfbf?source=rss-b9f4b02b86da------2"><img src="https://cdn-images-1.medium.com/max/2600/0*23O4L3BhWndVgIb0" width="6000"></a></p><p class="medium-feed-snippet">A century after Wittgenstein, AI is reshaping our language games</p><p class="medium-feed-link"><a href="https://medium.com/philosophytoday/do-large-language-models-understand-256e8a09dfbf?source=rss-b9f4b02b86da------2">Continue reading on Philosophy Today »</a></p></div>]]></description>
            <link>https://medium.com/philosophytoday/do-large-language-models-understand-256e8a09dfbf?source=rss-b9f4b02b86da------2</link>
            <guid isPermaLink="false">https://medium.com/p/256e8a09dfbf</guid>
            <category><![CDATA[language]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[philosophy]]></category>
            <dc:creator><![CDATA[Natasha Sachsenmeier]]></dc:creator>
            <pubDate>Fri, 07 Nov 2025 16:29:51 GMT</pubDate>
            <atom:updated>2025-11-10T18:53:08.558Z</atom:updated>
        </item>
    </channel>
</rss>