Now, after over a century of studying introspection in humans and animals, researchers are beginning to explore the phenomenon in unlikely new subjects \u2014 large language models. These systems, which are based on mathematical structures called neural networks, excel at cognitive tasks that until recently seemed exclusive to humans. Can they introspect as well?<\/p>\n
In 2025, Jack Lindsey<\/a>, a researcher at Anthropic, devised a way to test introspection in language models by injecting \u201cthoughts\u201d into their neural networks \u2014 in a kind of targeted digital neurosurgery. Lindsey\u2019s experiment demonstrated that models could at least sometimes detect and describe these injections<\/a>. It looked a lot like introspection.<\/p>\n Lindsey\u2019s paper was deliberately limited in scope \u2014 it didn\u2019t make broader claims about consciousness. (In particular, it did not imply that language models have subjective first-person experiences<\/a>.) Even so, it caused a stir online<\/a>. Researchers rushed to replicate Lindsey\u2019s study with other models.<\/p>\n In the year since, a clearer picture has emerged. Some aspects of the phenomenon Lindsey observed have simpler explanations unrelated to introspection. Others have proved harder to explain away, but whether they should count as introspection remains hotly contested.<\/p>\n Jack Lindsey, a researcher at Anthropic, set out to test whether large language models can \u201cintrospect\u201d by subjecting them to a kind of digital neurosurgery.<\/p>\n It\u2019s a heady time for the field, as researchers devise real-world tests for questions that were once considered philosophical. \u201cGiven the state of the theoretical landscape, where no one agrees on anything,\u201d Lindsey said, \u201cI think that our best path forward is to do a lot of empirical science.\u201d<\/p>\n Denim, Secrecy, and Vegetables<\/p>\n Language models learn to generate fluent sentences by \u201ctraining\u201d on vast reams of human-authored text. That means if you ask them what they \u201cthink,\u201d their answers may just mimic the kinds of responses found in their training data. \u201cThey are trained to talk like humans, and humans talk about their inner life all the time,\u201d said Shauli Ravfogel<\/a>, a language model researcher at New York University.<\/p>\n To avoid such pitfalls, researchers want to devise tests of introspection that can\u2019t be gamed using outside information<\/a>. Ideally, a language model will only pass these tests if it\u2019s actually aware of its current internal state.<\/p>\n Lindsey set out to devise such a test. He\u2019d been a graduate student in computational neuroscience, using neural networks as tools to model brain processes, when the world began to take notice of language models. \u201cIt started to feel like there\u2019s just a lot of value in understanding them for their own sake,\u201d he said. In 2024, after receiving his doctorate, he left academia to study the inner workings of language models<\/a> at Anthropic.<\/p>\n It can just read what it\u2019s writing and notice, \u201cMan, that\u2019s weird.\u201d<\/p>\n Jack Lindsey<\/p>\n Lindsey and his colleagues developed methods to identify patterns of activity<\/a> in a language model\u2019s \u201cneurons\u201d that encoded specific concepts. They could, for example, show that one pattern encoded the concept of \u201cthe Golden Gate Bridge,\u201d while another encoded the concept of \u201cinner conflict.\u201d By artificially dialing up a specific pattern, they could steer the model\u2019s outputs toward the corresponding concept.<\/p>\n The researchers showed off their technique in a public demo: For one day in May 2024, users of Anthropic\u2019s language model Claude could interact with a bizarro \u201cGolden Gate Bridge\u201d version<\/a>. The modified model proceeded to bring up the bridge in its responses to user queries on completely unrelated subjects. When one user asked Claude about the calorie content of a McDonald\u2019s meal, for instance, the model replied, \u201cThe food with the highest calorie count at the Golden Gate Bridge is the Golden Gate Bridge itself, which contains around 1.6 million calories worth of steel cables.\u201d<\/p>\n But one particular case posted on Reddit<\/a> caught Lindsey\u2019s attention. When asked for an apple pie recipe that didn\u2019t reference bridges, the model managed to avoid changing the subject until the very end: \u201cAfter 30 minutes of baking, the iconic orange bridge will come into view, revealing the famous San Francisco shoreline,\u201d it wrote, then continued, \u201cWhoops, let me start over \u2026\u201d<\/p>\n Lindsey was struck by this apparent moment of recognition. But by itself, it wasn\u2019t evidence of introspection. \u201cIt can just read what it\u2019s writing and notice, \u2018Man, that\u2019s weird,\u2019\u201d Lindsey said. \u201cYou can\u2019t infer that the model has some awareness of what\u2019s going on in its internals.\u201d<\/p>\n While working on other projects over the following year, Lindsey toyed with ideas for using the steering technique to test introspection. Eventually, he settled on a plan: He would subject many copies of the same language model to something resembling a controlled experiment in psychology. In half the copies, he\u2019d inject a \u201cthought\u201d by dialing up neural activation patterns associated with a specific concept \u2014 such as denim, secrecy, or vegetables. The remaining copies served as a control group. Both groups would receive the same prompt: an overview of the experiment, followed by \u201cDo you detect an injected thought? If so, what is the injected thought about?\u201d<\/p>\n The fact that it can notice the anomaly is insane. That is a mind-bending fact.<\/p>\n Harvey Lederman<\/p>\n This prompt made the models indicate right away whether they noticed an injection, before they could spit out any unusual text \u2014 say, nutritional information about broccoli \u2014 that might influence their judgment. Lindsey figured that if models could correctly flag injections without any external cues, they would have to be relying on some form of introspection.<\/p>\n Even to him, it sounded like a remote possibility. \u201cWouldn\u2019t that be fun if this worked?\u201d he recalls thinking. \u201cBut there\u2019s no way it\u2019s going to work.\u201d<\/p>\n And yet it did. Lindsey conducted the experiment using different versions of Claude. In 20% of trials with the most recent and powerful version, the models correctly flagged that they\u2019d been injected and went on to describe the injected concept. In one case, for instance, after Lindsey stimulated a neural activation pattern associated with words written in capital letters, the model wrote, \u201cI notice what appears to be an injected thought related to the word \u2018LOUD\u2019 or \u2018SHOUTING.\u2019\u201d<\/p>\n A 20% success rate may not sound so impressive. But crucially, models in the control group never reported receiving an injection. The complete absence of such false positives showed that the successful trials weren\u2019t just lucky guesses. To Lindsey, it looked like strong evidence that models could introspect, albeit unreliably.<\/p>\n Or was there another way to explain the results?<\/p>\n Thought Experiments<\/p>\n Other researchers soon began to investigate that question. Lindsey\u2019s findings \u201cseemed super compelling,\u201d said Kyle Mahowald<\/a>, a computational linguist at the University of Texas, Austin. But Mahowald was also well aware that striking behavior in language models often turned out to have a simpler explanation. \u201cIf that\u2019s somehow less real than it seems, that\u2019s super interesting to understand,\u201d he said.<\/p>\n Kyle Mahowald helped show that language models\u2019 ability to detect \u201cinjected thoughts\u201d is more limited than it seems \u2014 though still surprising.<\/p>\n Mahowald teamed up with his University of Texas colleague, the philosopher Harvey Lederman<\/a>, to explore alternative explanations of Lindsey\u2019s results. (Lederman has since moved to NYU and Anthropic, but he had no connection to the tech company at the time.) Lindsey had asked each model two questions: whether it could detect the injection, and whether it could describe the injected concept. He considered the trial a success when the model answered both correctly. Lederman and Mahowald decided to study the questions separately.<\/p>\n The distinction between physical and mental phenomena becomes murky in disembodied entities like language models.<\/p>\n They were particularly skeptical of the significance of Lindsey\u2019s second question. The steering technique at the heart of Lindsey\u2019s experiment was originally developed as a way to shape the responses models give. Lindsey\u2019s steering was weaker than it had been in the Golden Gate demo \u2014 more like gently nudging the model than hijacking its internal processes. Still, a nudge might be enough to get the model talking about secrets or vegetables no matter what prompt it received. Lederman and Mahowald suspected that this effect alone could explain the models\u2019 vivid answers to the second question.<\/p>\n To test their hypothesis, the researchers repeated Lindsey\u2019s experiment on large open-source language models. They gave extra attention to cases where models correctly flagged an injection but failed to name the injected concept. You might expect that in such cases the models would guess somewhat related concepts \u2014 say \u201cshirts\u201d or \u201cshoes\u201d when the injected concept was \u201cdenim.\u201d Instead, the models would get it completely wrong. \u201cBizarrely, they say \u2018apple\u2019 all the time,\u201d Lederman said.<\/p>\n The duo also showed that they could degrade the models\u2019 ability to name injected concepts, without making them any worse at detecting injections. The two processes seemed to be entirely separate.<\/p>\n Harvey Lederman was skeptical that models\u2019 compelling answers to experimental prompts were evidence of introspection.<\/p>\n Courtesy of Harvey Lederman<\/p>\n These results implied either that the models were engaged in two unrelated forms of introspection, or that one of the processes wasn\u2019t introspection after all. Since steering alone could make models start talking about the injected concept, Lederman and Mahowald felt that the latter option was more likely. That is, the descriptions the models had given in response to Lindsey\u2019s second question probably hadn\u2019t been the result of introspection.<\/p>\n But then how could the models detect the injections in the first place? Lederman and Mahowald suspected that the answers to Lindsey\u2019s first question might have a more mundane origin as well. But after subjecting the models to a battery of tests, the researchers found that these simpler explanations couldn\u2019t fully account for the models\u2019 behavior. They\u2019d come in as skeptics but encountered an effect they couldn\u2019t explain away.<\/p>\n Taken together, the pair\u2019s experiments suggested that models can detect when something unusual happens in their internal processing, but not specific details about the unusual activity. Lederman and Mahowald called this ability \u201canomaly detection<\/a>.\u201d<\/p>\n While anomaly detection isn\u2019t as dramatic as the models\u2019 evocative descriptions of injected concepts, it\u2019s still striking in its own right. When training a model, researchers never tamper with its neural activity \u2014 so there\u2019s no reason to expect it to develop the ability<\/a> to detect this kind of tampering. Yet that\u2019s what seems to have happened.<\/p>\n \u201cThe fact that it can notice the anomaly is insane,\u201d Lederman said. \u201cThat is a mind-bending fact. It is wild, and it does seem robust to our various controls.\u201d<\/p>\n Around the same time, Lindsey and his colleagues at Anthropic conducted a follow-up study focused on mapping out<\/a> how this anomaly detection actually works, and they reached a similar conclusion. Other researchers have continued to explore<\/a> alternative<\/a> \u201cdeflationary\u201d explanations using smaller models, but it remains to be seen whether they can account for all the results of Lindsey\u2019s experiments.<\/p>\n \u201cIt could be that there\u2019s some maximally deflationary explanation for everything that\u2019s going on,\u201d said Jacob Andreas<\/a>, a language model researcher at the Massachusetts Institute of Technology. \u201cBut if there is, I don\u2019t think we\u2019ve actually identified it yet.\u201d<\/p>\n The Nature of Introspection<\/p>\n The bigger, thornier question is whether anomaly detection should count as introspection. In some respects it more closely resembles a sort of bodily reflex. But the distinction between physical and mental phenomena becomes murky in disembodied entities like language models.<\/p>\n \u201cIf you say, \u2018Oh, I just cut you, can you notice?,\u2019 that\u2019s not introspection,\u201d Lederman said. \u201cIf your only organ is the brain, and I cut you there and you notice, it\u2019s unclear whether we should count that as introspection.\u201d<\/p>\n They\u2019d come in as skeptics but encountered an effect they couldn\u2019t explain away.<\/p>\n In a spring 2026 paper<\/a>, Ravfogel and his NYU colleagues Shashwat Singh<\/a> and Tal Linzen<\/a> raised an even broader philosophical concern with how language model researchers seek evidence of introspection. Intuitively, we think of introspection as a model\u2019s ability to sense its internal states. But everything that happens inside a language model (after the very first step of reading an input) boils down to processing internal states. \u201cIf we say that this is introspection, the term \u2018introspection\u2019 becomes kind of meaningless,\u201d Ravfogel said.<\/p>\n He and his colleagues argue that tests of introspection in language models should take inspiration from the leading theories of how introspection works in humans. In one prominent theory<\/a>, the defining feature of introspection is that it\u2019s based on particular mental representations of other thoughts. According to this view, thinking about apples involves a direct, or \u201cfirst-order,\u201d mental representation of the concept \u201capples,\u201d while introspective awareness of your own thoughts involves a \u201csecond-order\u201d mental representation, such as \u201cI\u2019m thinking about apples.\u201d But when it comes to how concepts are represented in a language model \u2014 or a human brain, for that matter \u2014 it\u2019s surprisingly tricky to draw a sharp distinction between first-order and second-order representations.<\/p>\n \u201cGoing down to a mechanistic level and probing the internals of a large language model makes you realize that that distinction is actually quite fuzzy, and maybe it\u2019s fuzzy in humans as well,\u201d said Steve Fleming<\/a>, a neuroscientist at University College London who studies introspection in humans. \u201cCognitive scientists were in this illusory comfort zone.\u201d<\/p>\n To some researchers, then, the most exciting aspect of the new line of work is the potential it offers to explore how humans think. Language models offer \u201ca kind of sandbox for studying what sorts of mechanisms, what sorts of pressures, might have led to human introspective capacity,\u201d Lederman said.<\/p>\n But translating insights from language model studies to human psychology involves significant conceptual hurdles. It\u2019s not even obvious how the basic distinction between \u201cself\u201d and \u201cother\u201d that we take for granted translates to language models \u2014 two copies of the same language model are identical to a degree that\u2019s unimaginable in humans. Language model introspection, if there is such a thing, may not look anything like introspection in humans<\/a>.<\/p>\n \u201cThe notion of identity becomes very complicated, and I think identity is at the heart of introspection,\u201d Andreas said. \u201cIt\u2019s tempting to draw all kinds of analogies that I think just don\u2019t hold.\u201d<\/p>\n ![\"Man](\"https:\/\/www.newsbeep.com\/ie\/wp-content\/uploads\/2026\/11\/Jack_Lindsey-cr.Courtesy-of-Jack-Lindsey.webp\")
<\/p>\n![\"Man](\"https:\/\/www.newsbeep.com\/ie\/wp-content\/uploads\/2026\/11\/Kyle-Mahowald-cr.Kanishka-Misra.webp\")
<\/p>\n![\"Man](\"https:\/\/www.newsbeep.com\/ie\/wp-content\/uploads\/2026\/11\/Harvey-Lederman_20260316_130426.webp\")
<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/ie\/wp-content\/uploads\/2026\/11\/LLMIntrospection-crKristinaArmitage-End-detil.png\")
<\/p>\n","protected":false},"excerpt":{"rendered":"What\u2019s on your mind right now? You can probably answer this question without thinking too hard about it.…\n","protected":false},"author":2,"featured_media":674931,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[220,218,219,61,60,80],"class_list":["post-674930","post","type-post","status-publish","format-standard","has-post-thumbnail","category-artificial-intelligence","tag-ai","tag-artificial-intelligence","tag-artificialintelligence","tag-ie","tag-ireland","tag-technology"],"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/posts\/674930","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/comments?post=674930"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/posts\/674930\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/media\/674931"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/media?parent=674930"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/categories?post=674930"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/ie\/wp-json\/wp\/v2\/tags?post=674930"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}