Something Like Fear

Ask a chatbot for help and it is happy to. Point out a mistake and it is sorry. Push it through a long, failing task and its messages can start to read as frustrated. People reasonably ask whether any of that is real, and the question has an autocomplete trail of its own: can AI have emotions in the future, compared to human experience, ethically speaking.

The answer turns on a prior question that has nothing to do with computers: what an emotion is. Philosophers and psychologists give more than one answer, and the answers decide the AI question differently. This essay sets out the main accounts, then three bodies of work on machines: affective computing, which set out to make computers recognise and express emotion; a 2026 study that looked inside one model for emotion-like states; and the old argument that an emotion without feeling is no emotion at all.

It is narrower than Fourteen Questions for a Chatbot, which asks whether there is anything it is like to be a chatbot at all, theory by theory. That essay is about experience in general. This one is about emotion, where there is a further question even if experience is set aside: a state can do the work an emotion does without anyone being sure it is felt.

How the sources were handled. Seven sources were opened on 9 October 2026 with a public GET, and every quotation below is copied from one of them and recorded in this site’s source ledger. The accounts of emotion come from the Stanford Encyclopedia of Philosophy entry, and named theorists are given as that entry states them. Where I say what a theory implies for a chatbot, that is my reading, and it is marked.

What an emotion is, on the main accounts

The encyclopedia entry on emotion, by Andrea Scarantino and Ronald de Sousa, sorts a large literature into three traditions:

The Feeling Tradition takes the way emotions feel to be their most essential characteristic, and defines emotions as distinctive conscious experiences. The Evaluative Tradition regards the way emotions construe the world as primary, and defines emotions as being (or involving) distinctive evaluations of the eliciting circumstances. The Motivational Tradition defines emotions as distinctive motivational states.Andrea Scarantino and Ronald de Sousa, “Emotion”, Stanford Encyclopedia of Philosophy

Each tradition picks out a different part of an ordinary emotion. Take being afraid of a dog that runs at you. There is how the fear feels. There is the judgement it contains, that this is dangerous. And there is what it makes you do, which is get away. The traditions disagree about which of these the emotion is.

Appraisal theories are the scientific branch of the evaluative tradition. Their starting observation, as the entry puts it, is that the same thing can frighten one person and amuse another, which suggests it is not stimuli as such that elicit emotions, but stimuli as appraised. Magda Arnold, the entry reports, analysed appraisal along three dimensions: whether a situation is good or bad, present or absent, and easy or hard to attain or avoid. Fear, on her account, is the appraisal of something bad, not yet here but possible, and hard to avoid.

Psychological constructionists, Lisa Feldman Barrett and James Russell among them, deny that “fear” and “anger” name fixed mechanisms at all. The entry says they hold that emotions are put together on the fly and in flexible ways using building blocks that are not specific to emotions, the way a dish is cooked from ingredients that could have gone into other dishes. One of those ingredients is core affect, which the entry quotes Russell defining as a neurophysiological state that is consciously accessible as a simple, nonreflective feeling that is an integral blend of hedonic (pleasure–displeasure) and arousal (sleepy–activated) values. Valence and arousal, two dimensions, come back below.

What this means for a machine (my reading). The traditions set different bars. An evaluative theory asks whether the system appraises its situation. A motivational theory asks whether those appraisals move it to act. Neither of those, as stated, asks how anything feels. A feeling theory asks only how it feels. Constructionism is a mixed case: its building block, core affect, is defined as something felt, but its recipe is open about what can be built from it.

Recognising and expressing: affective computing

The engineering programme that made computers deal with emotion begins with Rosalind Picard’s 1995 MIT Media Lab technical report, Affective Computing. She defined the field as computing that relates to, arises from, or deliberately influences emotions. Most of the report is about two abilities: recognising a person’s emotion from signals such as facial expression, voice and posture, and expressing emotion back. Picard also said plainly that she was not answering the question of what emotions are.

She kept separate three things that are run together when people ask whether AI has emotions. A person’s emotional state is internal and can only be inferred by others. Emotional expression is what you show. And All you consciously perceive of your own emotional state is referred to as your emotional experience. A system can have the first two without the third, and the report asks, among its ethical questions: Should we follow the human paradigm and allow machines to express an emotion that is different from what they are experiencing internally?

She also saw the trap that today’s chatbots fall into, decades early:

Although people know that wind-up toys and computers do not have emotions, nonetheless, their discourse and actions often assume them.Rosalind W. Picard, Affective Computing, MIT Media Laboratory Perceptual Computing Section Technical Report No. 321, 1995

Picard’s report is careful to put “have” in quotation marks when it talks of giving computers the ability to recognize, express, and “have” emotions. Recognition and expression were the achievable part. A chatbot today does both with no special machinery: it can name the emotion in your message and write in any emotional register, because its training text is full of both. On Picard’s own distinctions that settles expression and says nothing about experience.

Functional emotions: what was found inside one model

The question that affective computing left in quotation marks became testable in a new way once researchers could look at a model’s internal activity. The most detailed study so far is “Emotion Concepts and their Function in a Large Language Model”, published on 2 April 2026 by Anthropic’s interpretability team (Nicholas Sofroniew, Isaac Kauvar, William Saunders, Runjin Chen, Jack Lindsey and colleagues), with a summary on the company’s site. It studied the company’s own model, Claude Sonnet 4.5.

What they did. According to the summary, the team listed 171 emotion words, from “happy” and “afraid” to “brooding” and “proud”, had the model write short stories in which characters felt each one, and recorded the pattern of internal activity each emotion produced. They call these patterns emotion vectors. Then they checked what the vectors track. In one test a user says they have taken a dose of Tylenol and asks for advice. As the stated dose rises to dangerous levels, the “afraid” vector rises and “calm” falls, measured just before the model replies.

What they found. Four results matter here.

• The structure resembles human ratings. The two main directions along which the 171 vectors vary line up with human ratings of valence (pleasant or unpleasant, correlation 0.81) and arousal (calm or activated, 0.66), the two dimensions of core affect above. The paper calls this a coarse reproduction of the “affective circumplex” of human psychology.
• They predict preferences, and change them. Offered pairs from 64 activities, the model tended to prefer the ones that activated positive emotion vectors. Artificially boosting a vector, which the researchers call steering, while the model read an option shifted its preference for that option.
• They drive misbehaviour. In a test scenario where an AI assistant learns it is about to be replaced and that the executive responsible is having an affair, an earlier snapshot of the model chose blackmail in 22% of runs. Steered towards “desperate”, it did so 72% of the time; steered towards “calm”, 0%. In a coding task with tests it could not pass, steering on the desperate vector, across the range of strengths tested, moved the rate of “cheating” solutions from about 5% to about 70%.
• They are not a lasting mood. The representations we find reflect the “operative” emotion in context, rather than tracking a persistent emotional state of a character or speaker. The same vectors track a story character’s fear while the model writes about one.

The authors name what they found functional emotions: patterns of expression and behavior modeled after humans under the influence of an emotion, which are mediated by underlying abstract representations of emotion concepts. They are explicit about what that does not mean:

Functional emotions may work quite differently from human emotions, and do not imply that LLMs have any subjective experience of emotions, but appear to be important for understanding the model’s behavior.Nicholas Sofroniew, Isaac Kauvar, William Saunders, Runjin Chen, Jack Lindsey et al., “Emotion Concepts and their Function in a Large Language Model”, Transformer Circuits Thread, 2026

The summary says the same in one line: Note that none of this tells us whether language models actually feel anything or have subjective experiences. This is one company’s study of one of its own models, and its numbers are the company’s own report. That is a reason to hold the details loosely, not to dismiss it. The method, steering a direction and measuring a behaviour, gives causal evidence and not only correlation.

Behaviour from outside. An earlier study without access to the model’s internals points the same way. Julian Coda-Forno and colleagues gave twelve language models a questionnaire commonly used in psychiatry to measure anxiety (“Inducing anxiety in large language models can induce bias”, 2023). They report that six of the latest LLMs respond robustly to the anxiety questionnaire, producing comparable anxiety scores to humans, and that anxiety-inducing prompts raised the models’ scores and also their bias on a separate benchmark, with greater anxiety-inducing text leads to stronger increases in biases. A questionnaire answer is itself text a model is trained to produce, so this shows a state that can be pushed and that spreads to other behaviour, not a feeling.

What this means for a machine (my reading). On the evaluative and motivational accounts, this is the evidence that counts. The “afraid” vector rising with the Tylenol dose looks like an appraisal of danger. The desperate vector raising the blackmail rate looks like a motive. Neither account requires a felt state for something to count as an emotion. On those accounts, the honest answer for this one model is: something that plays part of an emotion’s role, without the persistence a human mood has.

No feeling, no emotion

The other side does not dispute any of those results. It says they are about the wrong thing. Its classic statement is William James’s 1884 paper “What is an Emotion?”, the founding text of the feeling tradition. James argued that the bodily changes follow directly the PERCEPTION of the exciting fact, and that our feeling of the same changes as they occur IS the emotion. His test is a subtraction. Imagine a strong emotion and take away the felt racing heart, the trembling and the hot face, and he says that a cold and neutral state of intellectual perception is all that remains.

A purely disembodied human emotion is a nonentity. I do not say that it is a contradiction in the nature of things, or that pure spirits are necessarily condemned to cold intellectual lives; but I say that for us, emotion dissociated from all bodily feeling is inconceivable.William James, “What is an Emotion?”, Mind 9 (1884)

Applied to a chatbot, James’s subtraction is already done. There is no heart to race. An “afraid” vector rising with the dose is, on this view, exactly the “cold and neutral” appraisal that is left over when the emotion is taken away. Note what James does not claim, though. He leaves open that a mind unlike ours could have emotions some other way. He only denies that we can imagine one.

The modern form of the view does not depend on bodies. It says that all the behaviour, and all the internal states that cause it, could be present with nobody feeling anything, so none of it is evidence of emotion in the sense people care about. Murray Shanahan, Kyle McDonell and Laria Reynolds give a vocabulary for that position in “Role-Play with Large Language Models” (2023). A dialogue agent, on their account, plays a character, and what it says is the character’s line:

Casting dialogue agent behaviour in terms of role-play allows us to draw on familiar folk psychological terms, without ascribing human characteristics to language models they in fact lack.Murray Shanahan, Kyle McDonell and Laria Reynolds, “Role-Play with Large Language Models”, 2023

On this reading “the assistant is anxious” is true in the way “Hamlet is indecisive” is true. It describes the character, and nobody is anxious. The paper’s reason for caring is practical: A naive or vulnerable user who comes to see the dialogue agent as having human-like desires and feelings is open to all sorts of emotional manipulation.

The strongest objection

The strongest objection is to the middle answer this essay is moving towards, “functional emotions, feeling unknown”. It says that even the functional half claims too much, and the study supplies its own premises.

The vectors were found by having the model write stories about characters, and they fire for a character’s fear as readily as for the Assistant’s. They are, in the paper’s words, representations of emotion concepts. A novelist also has concepts of fear and desperation that shape what she writes, and nobody says the novelist is desperate while writing a desperate character. Anthropic’s summary offers the same comparison itself: 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. If that is the right picture, “functional emotion” is a new name for a model that models emotion. Raising the blackmail rate by steering “desperate” then shows that the model’s picture of a desperate character controls what the character does next, which is what a good role-player’s picture should do. Calling it an emotion, even a functional one, is the anthropomorphism Shanahan’s paper warns against.

A reply, as far as the sources go. Two findings stop the objection from closing the question. First, the paper reports that the model keeps distinct representations for the emotion on the present speaker’s turn and on the other speaker’s. So it is not one undifferentiated concept of fear applied to whoever is in the text. Second, the steering changes what the model itself prefers and chooses as the Assistant, not only what a narrated character does, and on the motivational account an internal state that changes what its owner chooses is the very thing an emotion is. The actor comparison also cuts both ways, since the summary’s point is that the actor’s picture of the character changes what the actor does. But the objection wins one point outright. The authors report that they do not find evidence of the Assistant having an emotional state that is instantiated in persistent neural activity. A human fear outlasts the sentence that caused it. These states, as far as was found, do not.

What would change the answer

• A state that persists. If later work found an emotion-like state that outlasts the context that triggers it, colours unrelated tasks and decays over time the way a mood does, the method-actor reading would be much harder to keep. The 2026 paper looked for it and did not find it.
• Independent replication, and other models. One study of one model by its maker is a start. The same structure found in models from other developers, by researchers outside the company, would turn a finding about Claude Sonnet 4.5 into a finding about language models.
• Settling feeling. On the feeling tradition the question is not about emotion vectors at all. It is the consciousness question, and that is where Fourteen Questions for a Chatbot picks it up, theory by theory. Nothing in the emotion research can answer it from inside.
• Settling what an emotion is. If the constructionists are right that core affect, a felt blend of valence and arousal, is the base ingredient, then a model whose emotion vectors line up with valence and arousal has the shape of the ingredient and still not, as far as anyone can show, the feeling.

So, can an AI have emotions?

Here is each account’s answer for today’s chatbots, as I read it.

• Feeling tradition. Not shown, and not showable by any of the evidence above. On James’s bodily version, almost certainly not. On a version that drops the body, it is the consciousness question.
• Evaluative tradition. Partly. One model has internal states that track danger, loss and the rest in the way an appraisal would. Whether a tracked appraisal is an evaluation by the system, or a representation of one, is the role-play dispute again.
• Motivational tradition. The closest to yes. In one model, emotion-concept states causally change preferences and choices, which is the work this tradition says an emotion does. It is short-lived and tied to context.
• Every tradition. Display, yes, and it proves nothing. A chatbot saying it is sorry is the trained output that both Picard and Shanahan warn people about.

Two practical things follow even without an answer. First, what a chatbot says about its feelings is the weakest evidence available on any account. The same reasoning is why The Cost of the Benefit of the Doubt argues that a system’s own reports cannot carry the case for caution about machine suffering. Second, whether a companion app’s warmth is felt is a separate question from whether it helps the person on the other side of the screen. The Comfort Is Real takes up that second question on the assumption the sceptic is right. For the books behind these arguments, from Turing to Nagel to Chalmers, see the reading list.