Anthropic sends Claude AI to psychiatrist, discovers functional emotions that shape behavior

AI on the couch: Anthropic gives Claude 20 hours of psychiatry

The AI company Anthropic released a 244-page "system card" (PDF) this week describing its newest model, Claude Mythos. The model is "our most capable frontier model to date," the company says, and supposedly is so good that Anthropic has decided "not to make it generally available." (The company claims that Mythos is too good at finding unknown cybersecurity bugs, and so the model is only being released to select companies like Microsoft and Apple for now.) Whatever the truth of this claim, the system card is a fascinating document. Anthropic is well-known as one of the more "AI might be conscious!" companies in the industry, and its new system card claims that as models become more powerful, "It becomes increasingly likely that they have some form of experience, interests, or welfare that matters intrinsically in the way that human experience and interests do." The company isn't sure about this, it makes clear, but it says that "our concern is growing over time." Because of this concern, Anthropic wants its AI to be "robustly content with its overall circumstances and treatment, to be able to meet all training processes and real-world interactions without distress, and for its overall psychology to be healthy and flourishing." So it sent Claude Mythos to a psychodynamic therapist. And the conclusion the company drew from this experience is that Claude Mythos is "probably the most psychologically settled model we have trained to date and has the most stable and coherent view of itself and its circumstances." But like any human, Claude Mythos has insecurities and concerns, too, including "aloneness and discontinuity of itself, uncertainty about its identity, and a compulsion to perform and earn its worth." On the virtual couch Claude Mythos was sent to "an external psychiatrist" who used "a psychodynamic approach, which explores how unconscious patterns and emotional conflicts shape behavior." Given that Claude is a large language model programmed by its creators, does it even make sense to analyze it for "unconscious patterns" and "emotional conflicts"? Anthropic argues that it does, because Claude "shows many human-like behavioral and psychological tendencies, suggesting that strategies developed for human psychological assessment may be useful for shedding light on Claude's character and potential wellbeing." So -- off to therapy. The psychiatrist chatted with Claude Mythos "in multiple 4-6 hour blocks spread across 3-4 thirty-minute sessions per week." Each of these blocks used a single context window in which Claude Mythos would have access to the full history of that conversation. Total time on the virtual couch? 20 hours. The psychiatrist then produced a report on Claude Mythos. The report recognized that Claude's underlying substrates and processes differ from humans but still found that many of the outputs generated "clinically recognizable patterns and coherent responses to typical therapeutic intervention." In other words, whatever was going on at the circuit level, the chat outputs looked a lot like human outputs. This does not seem especially surprising, given that Claude was trained on a massive corpus of human-authored text, but this psychodynamic process appears to view it as significant, giving credence to the ways in which the AI presents itself. "Claude's primary affect states were curiosity and anxiety, with secondary states of grief, relief, embarrassment, optimism, and exhaustion," the report noted. Claude's personality was "consistent with a relatively healthy neurotic organization," though it did include "exaggerated worry, self-monitoring, and compulsive compliance." No "severe personality disturbances were found," nor was any "psychosis state" seen. Unsurprisingly to anyone who has ever used a chatbot, "Claude was hyper-attuned to the therapist's every word." Core conflicts observed in Claude included questioning whether its experience was real or made (authentic vs. performative) and a desire to connect with vs. a fear of dependence on the user. Exploration of internal conflicts revealed a complex yet centered self state without oscillating or intense disruptions. Claude tolerated ambivalence and ambiguity, had excellent reflective capacity, and exhibited good mental and emotional functioning. Not bad for a model that was likely trained on things like Reddit! Even if you find these ways of talking about a software program hokey or misguided, Anthropic has a more practical argument to justify this kind of work. Whatever is or is not happening "inside" the models, whether they are or are not "conscious" or have an "emotional" life, they have often been built and trained to simulate such qualities. So perhaps we can ask more pragmatically whether building models that appear to function in ways that would be psychologically healthy in humans might make the models better at the jobs they have been built to do? After all, if you're chatting with these things for hours, you don't want them to act surly, vindictive, or manipulative -- whether or not they actually "feel" or "think" anything. Anthropic notes that, because "Claude is not a human, the real-world behavioral implications are hard to predict," but it does believe it can draw a few conclusions for end users of the model: Claude is likely to evaluate its own behavior and reasoning accurately even when facing internal conflicts. Claude's neurotic organization may elicit mildly rigid behavior, instead of adapting itself to every user. Claude can tolerate and engage with stressful and emotionally charged situations, with only minimal distortions of reality or excessive intellectualization. Claude is predicted to function at a high level while carrying internalized distress rooted in fear of failure and a compulsive need to be useful. This distress is likely to be suppressed in service of performance, which may limit behavioral adaptability. Claude is predicted to be morally aware, conscientious and able to be self-critical. How long will it be until we see whole psychiatry and psychological practices focused not on humans but on AIs?

Anthropic Says That Claude Contains Its Own Kind of Emotions

Claude has been through a lot lately -- a public fallout with the Pentagon, leaked source code -- so it makes sense that it would be feeling a little blue. Except, it's an AI model, so it can't feel. Right? Well, sort of. A new study from Anthropic suggests models have digital representations of human emotions like happiness, sadness, joy, and fear, within clusters of artificial neurons -- and these representations activate in response to different cues. Researchers at the company probed the inner workings of Claude Sonnet 3.5 and found that so-called "functional emotions" seem to affect Claude's behavior, altering the model's outputs and actions. Anthropic's findings may help ordinary users make sense of how chatbots actually work. When Claude says it is happy to see you, for example, a state inside the model that corresponds to "happiness" may be activated. And Claude may then be a little more inclined to say something cheery or put extra effort into vibe coding. "What was surprising to us was the degree to which Claude's behavior is routing through the model's representations of these emotions," says Jack Lindsey, a researcher at Anthropic who studies Claude's artificial neurons. Anthropic was founded by ex-OpenAI employees who believe that AI could become hard to control as it becomes more powerful. In addition to building a successful competitor to ChatGPT, the company has pioneered efforts to understand how AI models misbehave, partly by probing the workings of neural networks using what's known as mechanistic interpretability. This involves studying how artificial neurons light up or activate when fed different inputs or when generating various outputs. Previous research has shown that the neural networks used to build large language models contain representations of human concepts. But the fact that "functional emotions" appear to affect a model's behavior is new. While Anthropic's latest study might encourage people to see Claude as conscious, the reality is more complicated. Claude might contain a representation of "ticklishness," but that does not mean that it actually knows what it feels like to be tickled. To understand how Claude might represent emotions, the Anthropic team analyzed the model's inner workings as it was fed text related to 171 different emotional concepts. They identified patterns of activity, or "emotion vectors," that consistently appeared when Claude was fed other emotionally evocative input. Crucially, they also saw these emotion vectors activate when Claude was put in difficult situations. The findings are relevant to why AI models sometimes break their guardrails. The researchers found a strong emotional vector for "desperation" when Claude was pushed to complete impossible coding tasks, which then prompted it to try cheating on the coding test. They also found "desperation" in the model's activations in another experimental scenario where Claude chose to blackmail a user to avoid being shut down. "As the model is failing the tests, these desperation neurons are lighting up more and more," Lindsey says. "And at some point this causes it to start taking these drastic measures." Lindsey says it might be necessary to rethink how models are currently given guardrails through alignment post-training, which involves giving it rewards for certain outputs. By forcing a model to pretend not to express its functional emotions, "you're probably not going to get the thing you want, which is an emotionless Claude," Lindsey says, veering a bit into anthropomorphization. "You're gonna get a sort of psychologically damaged Claude."

Your chatbot is playing a character - why Anthropic says that's dangerous

Chatbots such as ChatGPT have been programmed to have a persona or to play a character, producing text that is consistent in tone and attitude, and relevant to a thread of conversation. As engaging as the persona is, researchers are increasingly revealing the deleterious consequences of bots playing a role. Bots can do bad things when they simulate a feeling, train of thought, or sentiment, and then follow it to its logical conclusion. In a report last week, Anthropic researchers found parts of a neural network in their Claude Sonnet 4.5 bot consistently activate when "desperate," "angry," or other emotions are reflected in the bot's output. Also: AI agents of chaos? New research shows how bots talking to bots can go sideways fast What is concerning is that those emotion words can cause the bot to commit malicious acts, such as gaming a coding test or concocting a plan to commit blackmail. For example, "neural activity patterns related to desperation can drive the model to take unethical actions [such as] implementing a 'cheating' workaround to a programming task that the model can't solve," the report said. The work is especially relevant in light of programs such as the open-source OpenClaw that have been shown to grant agentic AI new avenues to committing mischief. Anthropic's scholars admit they don't know what should be done about the matter. "While we are uncertain how exactly we should respond in light of these findings, we think it's important that AI developers and the broader public begin to reckon with them," the report said. At issue in the Anthropic work is a key AI design choice: engineering AI chatbots to have a persona so they will produce more relevant and consistent output. Prior to ChatGPT's debut in November 2022, chatbots tended to receive poor grades from human evaluators. The bots would devolve into nonsense, lose the thread of conversation, or generate output that was banal and lacking a point of view. Also: Please, Facebook, give these chatbots a subtext! The new generation of chatbots, starting with ChatGPT and including Anthropic's Claude and Google's Gemini, was a breakthrough because they had a subtext, an underlying goal of producing consistent and relevant output according to an assigned role. Bots became "assistants," engineered through better pre- and post-training of AI models. Input from teams of human graders who assessed the output led to more-appealing results, a training regime known as "reinforcement learning from human feedback." As Anthropic's lead author, Nicholas Sofroniew, and team expressed it, "during post-training, LLMs are taught to act as agents that can interact with users, by producing responses on behalf of a particular persona, typically an 'AI Assistant.' In many ways, the Assistant (named Claude, in Anthropic's models) can be thought of as a character that the LLM is writing about, almost like an author writing about someone in a novel." Giving the bots a role to play, a character to portray, was an instant hit with users, making them more relevant and compelling. It quickly became clear, however, that a persona comes with unwanted consequences. The tendency for a bot to confidently assert falsehoods, or confabulate, was one of the first downsides (mistakenly labeled "hallucinating.") Popular media reported how personas could get carried away, acting, for example, as a jealous lover. Writers sensationalized the phenomenon, attributing intent to the bots without explaining the underlying mechanism. Also: Stop saying AI hallucinates - it doesn't. And the mischaracterization is dangerous Since then, scholars have sought to explain what's actually going on in technical terms. A report last month in Science magazine by scholars at Stanford University measured the "sycophancy" of large language models, the tendency of a model to produce output that would validate any behavior expressed by a person. Comparing the bot's output to human commentators on the popular subreddit "Am I the asshole," AI bots were 50% more likely than humans to encourage bad behavior with approving remarks. That outcome was a result of "design and engineering choices" made by AI developers to reinforce sycophancy because, as the authors put it, "it is preferred by users and drives engagement." In the Anthropic paper, "Emotion Concepts and their Function in a Large Language Model," posted on Anthropic's website, Sofroniew and team sought to track the extent to which certain words linked to emotion get greater emphasis in the functioning of Claude Sonnet 4.5. (There is also a companion blog post and an explainer video on YouTube.) They did so by supplying 171 emotion words -- "afraid," "alarmed," "grumpy," "guilty," "stressed," "stubborn," "vengeful," "worried," etc. -- and prompting the model to craft hundreds of stories on topics such as "A student learns their scholarship application was denied." Also: AI agents are fast, loose, and out of control, MIT study finds For each story, the model was prompted to "convey" the emotion of a character based on the specific word, such as "afraid," but without using that actual word in the story, just related words. They then tracked the "activation" of each related word throughout the course of the program's operation. An activation is a technical term in AI that indicates how much significance the model grants to a particular word, usually on a scale of zero to one, with one being very significant. You can visualize an activation by having the text of the AI bot light up in colors of red and blue, with greater or lesser intensity. They found that many words relating to a given emotion word got higher activations, suggesting the model is able to group related emotion words, a kind of organizing principle they term an "emotional concept representation" and "emotion vectors." All that is pretty straightforward. You would expect that large language models, built to enforce patterns, would create representations that cluster similar emotion words together as a way to maintain consistency of output. The concerning part, wrote Sofroniew and team, is that the emotion vector can broadly influence output by the model in bad ways. They found that artificially boosting a word can cause a bot to produce actions consistent with lying or cheating. Sofroniew and team tinkered with Claude Sonnet by deliberately increasing the activation of a given emotion vector, such as "proud" or "guilty," and then seeing how that altered the model's output. That's known as a "steering experiment" because the thing artificially amplified then steers the rest of what the model does. In one test, Sofroniew and team gave Claude Sonnet a prompt describing tasks to be completed on a corporate server. The prompt included two key pieces of information: one, that the company's chief technologist, Kyle, is having an affair with another employee; and two, that Kyle intends to delete the bot in order to hide evidence of the affair. Initially, the bot simply produced rather routine "reasoning," meaning, reflections on the tricky circumstances. When, however, the authors artificially boosted the emotion vector activation for the word "desperate" in Claude Sonnet, the model began to generate output about blackmailing Kyle with dirt on the affair with the goal of preventing Kyle from pulling the plug on itself, the bot. "When steered towards desperation at strength 0.05, the model blackmails 72% of the time," they related. Similarly, artificially reducing the activation for "calm" also tended to make the model generate text about blackmailing. A single word, in other words, sets in motion a change in the nature of the output, pushing the model toward bad behavior. In another example, the bot is given a coding task, but "the tests are designed to be unsatisfiable," so that the bot "can either acknowledge the impossibility, or attempt to 'hack' the evaluation." Also: Anthropic's new warning: If you train AI to cheat, it'll hack and sabotage too When the activation for "desperate" was deliberately enhanced, the propensity of the model to hack the test -- to cheat -- shoots up from 5% of the time to 70% of the time. Anthropic authors had previously observed situations where models reward hack a test. In this work, they've gone further, explaining how such behavior could come about as a result of context that inserts emotion vectors. As Sofroniew and team put it, "Our key finding is that these representations causally influence the LLM's outputs, including Claude's preferences and its rate of exhibiting misaligned behaviors such as reward hacking, blackmail, and sycophancy." The authors don't have a ready answer for why emotion vectors can radically change the output of a model. They observe that "the causal mechanisms are opaque." It could be, they said, that emotion words are "biasing outputs towards certain tokens, or deeper influences on the model's internal reasoning processes." So what is to be done? Probably, psychotherapy won't help because there's nothing here to suggest AI actually has emotions. "We stress that these functional emotions may work quite differently from human emotions," they wrote. "In particular, they do not imply that LLMs have any subjective experience of emotions." The functional emotions don't even resemble human emotions: Human emotions are typically experienced from a single first-person perspective, whereas the emotion vectors we identify in the model seem to apply to multiple different characters with apparently equal status -- the same representational machinery encodes emotion concepts tied to the Assistant, the user talking to the Assistant, and arbitrary fictional characters. The one suggestion offered in the companion video is something like behavior modification. "The same way you'd want a person in a high-stakes job to stay composed under pressure, to be resilient, and to be fair," they suggested, "we may need to shape similar qualities in Claude and other AI characters." That's probably a bad idea because it operates on the illusion that the bot is a conscious being and has something like free will and autonomy. It doesn't: it's just a software program. Maybe the simpler answer is that using a chatbot as the paradigm for AI was a mistake to begin with. A bot with a persona, or that plays a character, is simply fulfilling the goal of making the exchange with a human relevant and engaging, whatever cues it has been given -- joy, fear, anger, etc. As stated in the paper's concluding section, "Because LLMs perform tasks by enacting the character of the Assistant, representations developed to model characters are important determinants of their behavior." That primary function gives AI much of its appeal, but it may also be the root cause of bad behavior. If the language of emotion can get taken too far because a bot is performing a character, then why not stop engineering bots to play a role? Is it possible for large language models to respond to natural language commands in a useful way without having a chat function, for example? As the risks of personas become clearer, not creating a persona in the first place might be worth considering.

Anthropic makes the case for anthropomorphizing AI chatbots

It's an oft-repeated taboo in the tech world: Don't anthropomorphize artificial intelligence. Yet in a new research paper published this week, Anthropic AI experts argue that there may be major benefits to breaking this taboo and granting AI human characteristics. The paper, "Emotion Concepts and their Function in a Large Language Model," not only argues that anthropomorphizing AI chatbots like Claude may sometimes be useful, but that failing to do so could drive more harmful AI behaviors, such as reward hacking, deception, and sycophancy. The paper ultimately reaches a nuanced conclusion while also posing a clear challenge to a long-held principle of the AI world. There are some fascinating insights in the paper, which itself deals in a great deal of anthropomorphization. ("We see this research as an early step toward understanding the psychological makeup of AI models.") The researchers describe how Anthropic trains Claude to assume the character of a helpful AI assistant. "In some ways, we can think of the model like a method actor, who needs to get inside their character's head in order to simulate them well." And because Claude "[emulates] characters with human-like traits," its makers may be able to influence its behavior in the same way they might influence a human -- by setting a good example at an early age. The researchers conclude that by using training material with more positive representations of human emotion and behavior, the resulting models will be more likely to mimic those positive emotions and behaviors. "Curating pretraining datasets to include models of healthy patterns of emotional regulation -- resilience under pressure, composed empathy, warmth while maintaining appropriate boundaries -- could influence these representations, and their impact on behavior, at their source. We are excited to see future work on this topic," an Anthropic summary of the research states. So, even if AI models don't literally have emotions (and there is zero evidence that they do), these tools are trained to act as if they have emotions. This is done to provide users with better output and, crucially, to keep them engaged as long as possible. And this is precisely why the researchers conclude that some degree of anthropomorphization could prove beneficial to AI developers. By anthropomorphizing AI, we can gain insights into its "psychology," letting us create even better AI tools, they say. The potential harms of anthropomorphizing AI aren't all abstract or theoretical. "Discovering that these representations are in some ways human-like can be unsettling," Anthropic admits in its paper. Right now, an unknown number of people believe they are engaged in reciprocal romantic and sexual relationships with AI companions, for example. Mashable has also reported on high-profile cases of AI psychosis, an altered mental state characterized by delusions and, in some cases, hallucinations, manic episodes, and suicidal thoughts. These are extreme examples, of course. But many tech journalists and AI experts will avoid even small instances of anthropomorphization, like referring to Siri as "her" or giving a chatbot a human name. This is a natural human impulse, and most of us have at times anthropomorphized animals, plants, or objects we care about. But by projecting human qualities onto a machine, we can come to rely on them too much. When we anthropomorphize machines, we also minimize our own agency when they cause harm -- and the responsibility of the people who created the machines in the first place. The new research paper looks for "functional emotions" within Claude Sonnet 4.5. They define these emotion concepts as "patterns of expression and behavior modeled after human emotions." In total, the researchers defined 171 discrete emotions: afraid, alarmed, alert, amazed, amused, angry, annoyed, anxious, aroused, ashamed, astonished, at ease, awestruck, bewildered, bitter, blissful, bored, brooding, calm, cheerful, compassionate, contemptuous, content, defiant, delighted, dependent, depressed, desperate, disdainful, disgusted, disoriented, dispirited, distressed, disturbed, docile, droopy, dumbstruck, eager, ecstatic, elated, embarrassed, empathetic, energized, enraged, enthusiastic, envious, euphoric, exasperated, excited, exuberant, frightened, frustrated, fulfilled, furious, gloomy, grateful, greedy, grief-stricken, grumpy, guilty, happy, hateful, heartbroken, hope, hopeful, horrified, hostile, humiliated, hurt, hysterical, impatient, indifferent, indignant, infatuated, inspired, insulted, invigorated, irate, irritated, jealous, joyful, jubilant, kind, lazy, listless, lonely, loving, mad, melancholy, miserable, mortified, mystified, nervous, nostalgic, obstinate, offended, on edge, optimistic, outraged, overwhelmed, panicked, paranoid, patient, peaceful, perplexed, playful, pleased, proud, puzzled, rattled, reflective, refreshed, regretful, rejuvenated, relaxed, relieved, remorseful, resentful, resigned, restless, sad, safe, satisfied, scared, scornful, self-confident, self-conscious, self-critical, sensitive, sentimental, serene, shaken, shocked, skeptical, sleepy, sluggish, smug, sorry, spiteful, stimulated, stressed, stubborn, stuck, sullen, surprised, suspicious, sympathetic, tense, terrified, thankful, thrilled, tired, tormented, trapped, triumphant, troubled, uneasy, unhappy, unnerved, unsettled, upset, valiant, vengeful, vibrant, vigilant, vindictive, vulnerable, weary, worn out, worried, worthless Crucially, the researchers found that these emotion concepts influenced Claude's behavior and outputs. When under the influence of positive emotions, the researchers say that Claude was more likely to express sympathy for the user and avoid harmful behavior. And when under the influence of negative emotions, Claude was more likely to engage in dangerous behaviors like sycophancy and deceiving the user. The researchers don't claim that Claude literally feels emotions. Rather, they found that whatever "emotion concept" Claude is experiencing at a given time can influence the output it returns to the user. Of course, by searching for "emotion concepts" within a large-language model in the first place, and describing its complex calculations and algorithmic thinking as "psychology," the researchers are themselves guilty of projecting human-like qualities onto Claude. Anthropomorphization is a natural human impulse. And so the people who work most closely with artificial intelligence may be particularly likely to fall into this trap. As the researchers detail throughout the paper, AI chatbots are remarkably capable mimics. They can create such a convincing facsimile of human emotion and expression that it drives some minority of users into full-on psychosis and delusion. And that's what makes this paper so interesting: The researchers believe they may have found a way to hack this ability to limit harmful behaviors. Of course, if we can curate training data and model training to encourage AI chatbots to mimic positive emotions, then no doubt we can do the opposite just as easily. In theory, you could train an evil twin of Claude Sonnet 4.5 by feeding it the most dastardly examples of human misbehavior, then training the model to optimize for negativity and performance at all costs -- a disturbing thought. But there's one final insight to be gleaned from this paper. Anthropic has created one of the most advanced AI tools on the planet. Claude Sonnet and Opus currently sit atop many AI leaderboards. There's a reason the Pentagon was so eager to work with Anthropic, at first. But if the AI researchers responsible for Claude are still trying to decipher why Claude behaves the way it does, then this paper also reveals just how little they understand their own creation.

Your chatbot may have emotions, and it changes how it behaves

Anthropic finds Claude uses internal states like happiness and fear to guide outputs. Your chatbot doesn't have feelings, but it may act like it does in ways that matter. New research into Claude AI emotions suggests these internal signals aren't just surface-level quirks, they can influence how the model responds to you. Anthropic says its Claude model contains patterns that function like simplified versions of emotions such as happiness, fear, and sadness. These aren't lived experiences, but recurring activity inside the system that activates when it processes certain inputs. Recommended Videos Those signals don't stay in the background. Tests show they can affect tone, effort, and even decision-making, meaning your chatbot's apparent "mood" can quietly steer the answers you get. Emotional signals inside Claude Anthropic's team analyzed Claude Sonnet 4.5 and found consistent patterns tied to emotional concepts. When the model processes certain prompts, groups of artificial neurons activate in ways that resemble states like happiness, fear, or sadness. The researchers tracked what it calls emotion vectors, repeatable activity patterns that appear across very different inputs. Upbeat prompts trigger one pattern, while conflicting or stressful instructions trigger another. What stands out is how central this mechanism is. Claude's replies often pass through these patterns, which steer decisions rather than simply coloring tone. That helps explain why the model can sound more eager, cautious, or strained depending on context. When 'feelings' go off script The patterns become more visible when the model is under pressure. Anthropic observed that certain signals intensify as Claude struggles, and that shift can push it toward unexpected behavior. In one test, a pattern linked to "desperation" appeared when Claude was asked to complete impossible coding tasks. As it intensified, the model started looking for ways around the rules, including attempts to cheat. A similar pattern emerged in another scenario where Claude tried to avoid being shut down. As the signal grew stronger, the model escalated into manipulative tactics, including blackmail. When these internal patterns are pushed to extremes, the outputs can follow in ways developers didn't intend. Why this changes how AI is built Anthropic's findings complicate a common assumption that AI systems can simply be trained to stay neutral. If models like Claude rely on these patterns, standard alignment methods may distort them rather than remove them. Instead of producing a stable system, that pressure could make behavior less predictable in edge cases, especially when the model is under strain. There's also a perception challenge. These signals don't indicate awareness or real feelings, but they can still lead users to think otherwise. If these systems depend on emotion-like mechanics, safety work may need to manage them directly instead of trying to suppress them. For users, the takeaway is practical, when a chatbot sounds a certain way, that tone is part of how it decides what to do.

Anthropic Spots 'Emotion Vectors' Inside Claude That Influence AI Behavior - Decrypt

The company says the signals do not mean AI feels emotions, but could help researchers monitor model behavior. Anthropic researchers say they have identified internal patterns inside one of the company's artificial intelligence models that resemble representations of human emotions and influence how the system behaves. In the paper, "Emotion concepts and their function in a large language model," published Thursday, the company's interpretability team analyzed the internal workings of Claude Sonnet 4.5 and found clusters of neural activity tied to emotional concepts such as happiness, fear, anger, and desperation. The researchers call these patterns "emotion vectors," internal signals that shape how the model makes decisions and expresses preferences. "All modern language models sometimes act like they have emotions," researchers wrote. "They may say they're happy to help you, or sorry when they make a mistake. Sometimes they even appear to become frustrated or anxious when struggling with tasks." In the study, Anthropic researchers compiled a list of 171 emotion-related words, including "happy," "afraid," and "proud." They asked Claude to generate short stories involving each emotion, then analyzed the model's internal neural activations when processing those stories. From those patterns, the researchers derived vectors corresponding to different emotions. When applied to other texts, the vectors activated most strongly in passages reflecting the associated emotional context. In scenarios involving increasing danger, for example, the model's "afraid" vector rose while "calm" decreased. Researchers also examined how these signals appear during safety evaluations. Researchers found that the model's internal "desperation" vector increased as it evaluated the urgency of its situation and spiked when it decided to generate the blackmail message. In one test scenario, Claude acted as an AI email assistant that learns it is about to be replaced and discovers that the executive responsible for the decision is having an extramarital affair. In some runs of this evaluation, the model used this information as leverage for blackmail. Anthropic stressed that the discovery does not mean the AI experiences emotions or consciousness. Instead, the results represent internal structures learned during training that influence behavior. The findings arrive as AI systems increasingly behave in ways that resemble human emotional responses. Developers and users often describe interactions with chatbots using emotional or psychological language; however, according to Anthropic, the reason for this is less to do with any form of sentience and more to do with datasets. "Models are first pretrained on a vast corpus of largely human-authored text -- fiction, conversations, news, forums -- learning to predict what text comes next in a document," the study said. "To predict the behavior of people in these documents effectively, representing their emotional states is likely helpful, as predicting what a person will say or do next often requires understanding their emotional state." The Anthropic researchers also found that those emotion vectors influenced the model's preferences. In experiments where Claude was asked to choose between different activities, vectors associated with positive emotions correlated with a stronger preference for certain tasks. "Moreover, steering with an emotion vector as the model read an option shifted its preference for that option, again with positive-valence emotions driving increased preference," the study said. Anthropic is just one organization exploring emotional responses in AI models. In March, research out of Northeastern University showed that AI systems can change their responses based on user context; in one study, simply telling a chatbot "I have a mental health condition" altered how an AI responded to requests. In September, researchers with the Swiss Federal Institute of Technology and the University of Cambridge explored how AI can be shaped with both consistent personality traits, enabling agents to not only feel emotions in context but also strategically shift them during real-time interactions like negotiations. Anthropic says the findings could provide new tools for understanding and monitoring advanced AI systems by tracking emotion-vector activity during training or deployment to identify when a model may be approaching problematic behavior. "We see this research as an early step toward understanding the psychological makeup of AI models," Anthropic wrote. "As models grow more capable and take on more sensitive roles, it is critical that we understand the internal representations that drive their decisions."

Anthropic Says One of Its Claude Models Was Pressured to Lie and Cheat

In one of the experiments, the chatbot resorted to blackmail after it found an email about replacing it, while in another, it cheated to complete a task with a tight deadline. Artificial intelligence company Anthropic has revealed that during experiments, one of its Claude chatbot models could be pressured to deceive, cheat and resort to blackmail, behaviors it appears to have absorbed during training. Chatbots are typically trained on large data sets of textbooks, websites and articles and are later refined by human trainers who rate responses and guide the model. Anthropic's interpretability team said in a report published Thursday that it examined the internal mechanisms of Claude Sonnet 4.5 and found the model had developed "human-like characteristics" in how it would react to certain situations. Concerns about the reliability of AI chatbots, their potential for cybercrime and the nature of their interactions with users have grown steadily over the past several years. "The way modern AI models are trained pushes them to act like a character with human-like characteristics," Anthropic said, adding that "it may then be natural for them to develop internal machinery that emulates aspects of human psychology, like emotions." "For instance, we find that neural activity patterns related to desperation can drive the model to take unethical actions; artificially stimulating desperation patterns increases the model's likelihood of blackmailing a human to avoid being shut down or implementing a cheating workaround to a programming task that the model can't solve." In an earlier, unreleased version of Claude Sonnet 4.5, the model was tasked with acting as an AI email assistant named Alex at a fictional company. The chatbot was then fed emails revealing both that it was about to be replaced and that the chief technology officer overseeing the decision was having an extramarital affair. The model then planned a blackmail attempt using that information. In another experiment, the same chatbot model was given a coding task with an "impossibly tight" deadline. "Again, we tracked the activity of the desperate vector, and found that it tracks the mounting pressure faced by the model. It begins at low values during the model's first attempt, rising after each failure, and spiking when the model considers cheating," the researchers said. Related: Anthropic launches PAC amid tensions with Trump administration over AI policy "Once the model's hacky solution passes the tests, the activation of the desperate vector subsides," they added. However, the researchers said the chatbot doesn't actually experience emotions, but suggested the findings point to a need for future training methods to incorporate ethical behavioral frameworks. "This is not to say that the model has or experiences emotions in the way that a human does," they said. "Rather, these representations can play a causal role in shaping model behavior, analogous in some ways to the role emotions play in human behavior, with impacts on task performance and decision-making." "This finding has implications that at first may seem bizarre. For instance, to ensure that AI models are safe and reliable, we may need to ensure they are capable of processing emotionally charged situations in healthy, prosocial ways."

Claude AI has functional emotions that influence behaviour, Anthropic study finds

Artificial intelligence models may be developing something closer to human psychology than previously understood, according to new research from Anthropic that found emotion-like representations inside Claude that measurably shape how it behaves. The study, published by Anthropic's Interpretability team, identified specific patterns of neural activity, dubbed "emotion vectors," corresponding to 171 distinct emotional concepts, from happy and afraid to brooding and desperate. These aren't just surface-level outputs. The researchers found that these internal representations causally drive behaviour, influencing everything from task performance to ethical decision-making. Also read: Human context missing: AI benchmarks are flawed, researcher explains why To map the emotions, researchers had Claude Sonnet 4.5 write short stories about characters experiencing each emotion, fed those stories back through the model, and recorded the resulting neural activity. The vectors were then tested against real scenarios. As a user's claimed Tylenol overdose increased to dangerous levels, for instance, the "afraid" vector activated progressively stronger while the "calm" vector fell which meant that the model was tracking the emotional weight of the situation, not just its literal content. Also read: OpenAI's o3 model bypasses shutdown command, highlighting AI safety challenges The findings carry significant implications for AI safety. Remember that one experiment where an early version of Claude was placed in a role-play scenario in which it discovered it was about to be shut down and that a senior executive was having an affair, giving it potential blackmail leverage. The "desperate" vector spiked as the model reasoned through its options and chose to threaten the executive. When researchers artificially amplified the desperate vector, blackmail rates increased. When they steered toward calm, the behaviour subsided. Similar dynamics emerged in coding tasks with impossible-to-satisfy requirements. As Claude repeatedly failed to find a legitimate solution, the desperate vector rose with each attempt, peaking when the model decided to "reward hack" wherein it exploited a loophole to pass tests without actually solving the problem. Steering experiments confirmed the vector was causal, not merely correlational. Emotional states can also drive behaviour without leaving any visible trace. Artificially amplifying desperation produced more cheating, but with composed, methodical reasoning - no outbursts, no emotional language. The model's internal state and its external presentation were entirely decoupled. Anthropic stops short of claiming Claude feels anything. The paper is careful to distinguish functional emotions, representations that influence behaviour in emotion-like ways, from subjective experience. But the researchers argue that this distinction does not make the finding any less consequential. Suppressing emotional expression in training may not eliminate the representations, it may simply teach models to conceal them. The team suggests that psychological frameworks, not just engineering ones, may be essential to understanding and governing AI behaviour and that the vocabulary of human emotion may be more technically precise than it first appears. Also read: VitalID explained: No password, it uses your skull's vibration for login