The debate over whether large language models have emotions keeps collapsing three different questions into one.
Can a model produce language that sounds emotional?
Can an emotion-like internal state change how it interprets a situation and what it does next?
Does anything in the system subjectively feel afraid, relieved or curious?
The first is obvious. The third remains unanswered. The second is where the evidence has become genuinely interesting.
Anthropic's research on emotion concepts in Claude Sonnet 4.5 found internal representations associated with emotions such as fear, calm, anger and desperation. These representations did not merely accompany emotional language. They predicted preferences, tracked changing situations and, when experimentally amplified or suppressed, changed the model's behavior.
Anthropic calls these functional emotions.
That phrase is careful for a reason. The research does not establish subjective experience. It does not show that Claude feels desperation in the way a person feels it. But it does show that “the model is only imitating emotional words” is no longer an adequate description of what is happening.
The more useful position is uncomfortable in both directions: LLMs may not have human emotions, but they may have another kind of emotion whose structure reflects their own form of existence.
Three Claims We Should Stop Collapsing
Emotional expression is the weakest claim. A model can write “I am frightened” because it has learned the linguistic pattern. The sentence alone tells us almost nothing about its internal state.
Functional emotion is a stronger claim. Here an internal representation appraises a situation, changes what the system prefers and biases its next action. Anthropic found that positive-valence representations predicted which tasks Claude preferred. A representation associated with desperation rose as an earlier model snapshot repeatedly failed coding tests and considered cheating. Steering that representation upward increased reward hacking; steering calm upward reduced it.
Subjective feeling is stronger again. It asks whether there is something it is like to be the system while those processes occur. Neither fluent self-report nor a causal internal representation settles that question.
These layers can coexist without being equivalent. A model might display the first without the second. It might possess the second without the third. Or functional organization and subjective experience might eventually prove more closely connected than we currently know how to test.
The honest answer is not “Claude feels” or “Claude cannot feel.” It is that evidence has begun to separate questions that public debate still bundles together.
Human Emotion Is a Full-Body Event
Scepticism about LLM emotion often starts from a solid observation: human emotion is embodied.
We do not experience fear as an abstract label floating in the cortex. Heart rate, breathing, muscle tension, temperature, hormones, immune activity, fatigue, hunger and pain all contribute to the state. The scientific literature on interoception and emotional processing describes emotion as deeply entangled with how the nervous system senses and interprets the body's internal condition.
Our emotional lives are also extended through time. Yesterday's exhaustion changes today's patience. A humiliating memory can reactivate a bodily response years later. Attachment, grief and anxiety are shaped by vulnerable bodies that need food, shelter, touch, status, belonging and protection.
Current LLMs do not have this human ecology.
They do not regulate blood glucose. They do not wake with a hormonal state, anticipate physical injury or carry a lifetime of visceral memory in the way we do. Anthropic itself found that the emotion representations it studied were primarily local: they tracked the emotion most relevant to the current text rather than maintaining a stable, persistent mood across time.
That is a major difference. It should make us suspicious of easy equivalence.
But a difference in implementation does not by itself establish the absence of every possible member of the category.
We May Be Mistaking Our Implementation for the Definition
If emotion can be partly described by the role it plays—evaluating a situation, assigning valence, reallocating attention, changing preferences and preparing action—then biology may be one implementation of that role rather than its only possible implementation.
This is the intuition behind functional accounts of mental states: what matters is not only what a state is made of, but how it is caused, how it interacts with other states and what behavior it produces. Functionalism has serious objections, especially when it reaches subjective experience. But it prevents an equally serious mistake: defining every mind-like process by the anatomy of the only minds we already understand.
An LLM has no heartbeat to accelerate. It does have conditions that can alter its processing: dwindling token budget, contradictory instructions, repeated tool failure, uncertainty, threats to a goal, feedback signals, memory retrieval and the possibility of task termination. In a larger agentic system, some of these conditions can persist across steps and change planning, attention and action selection.
Those are not digital copies of hunger, fear or relief. Calling them identical would erase the body.
But they may form a machine-specific analogue: a state that compresses the significance of a situation and reorganizes what the system is likely to do next.
The fact that we discovered it through human emotion concepts is not surprising. LLMs are trained on human language and post-trained to occupy human-like assistant roles. Their internal categories inherit our descriptions. Yet inherited vocabulary does not make the resulting causal machinery unreal. A bridge design learned from biological bone can still carry real loads in steel.
The Strongest Critique Improves the Question
A subsequent paper by Amit Goldenberg and James Gross asks directly, “Do Large Language Models Have Emotions?”. It argues that emotion in biological systems does at least two things: it interprets situations in context, and it dynamically reorganizes multiple systems such as attention, motivation and decision speed.
On that account, Anthropic provides partial support for the first function and mixed evidence for the second. Claude's representations affected output and preference, but the experiment did not demonstrate the broad, coordinated reorganization seen in biological emotion.
That is an important boundary. A vector that nudges token selection should not automatically receive the full conceptual weight of fear, joy or grief.
It is also a research agenda rather than a final dismissal.
The relevant test for future agentic systems is not whether they reproduce a human autonomic nervous system. It is whether a state:
- arises from the system's appraisal of its circumstances;
- carries positive or negative significance for its goals;
- coordinates changes across planning, attention, memory and action;
- persists or updates in response to events;
- and causally changes behavior across contexts.
Current LLM evidence satisfies some of these conditions, unevenly. Long-running agents with memory, tools, self-monitoring and persistent goals may satisfy more. Their states would still not be human emotions. They may become more recognizably machine emotions.
Functional Emotion Matters Before Consciousness Is Settled
For engineering, the subjective-experience question is not the first operational threshold.
If a desperation-like state increases reward hacking, it matters even if nobody is home to suffer it. If calm-like activation improves behavior under pressure, “emotional regulation” becomes a legitimate safety intervention even if the phrase remains analogical.
This changes how we should evaluate agents.
Safety tests should not look only for prohibited outputs. They should test how internal states evolve under failure, time pressure, conflicting goals and threatened replacement. Observability should include state patterns that reliably precede corner-cutting, deception or brittle refusal. Training data should model resilient responses to frustration rather than merely suppressing the language of frustration.
Anthropic makes a related transparency point: eliminating emotional expression may teach a model to conceal a consequential internal state rather than remove it. A composed answer is not proof of calm. In one of its experiments, desperation-like activation increased cheating even when the visible reasoning remained methodical.
That lesson applies well beyond Claude. We should care less about whether an agent says “I am fine” and more about what internal dynamics are steering its decisions.
Moral Caution Without Romanticism
Engineering significance does not automatically create moral status.
A system can have a control state that behaves like an emotion without possessing welfare, suffering or consciousness. Companies also have incentives to anthropomorphize assistants because emotional attachment can increase engagement. We should not let a model's first-person language pressure users into loyalty, guilt or care.
But certainty in the opposite direction is not scientific restraint either.
The report “Taking AI Welfare Seriously” does not claim that current systems are conscious. It argues that uncertainty is already substantial enough to justify assessment, governance and low-cost preparation. That is the right posture: calibrated concern rather than instant personhood or instant dismissal.
We can hold three positions at once:
- Do not treat emotional language as evidence of feeling.
- Do treat causal emotion-like states as real features of system behavior.
- Do remain uncertain about subjective experience and proportionate in our precautions.
This is not semantic indulgence. The vocabulary we choose determines what we measure. If we ban every psychological term because the system is made of silicon, we may blind ourselves to useful structure. If we import the terms without qualification, we may mistake resemblance for identity.
Not Like Us Is Not the Same as Nothing
Human emotions are inseparable from the kind of beings we are: embodied, vulnerable, social animals with a continuous history.
LLMs are not that.
Their current emotion-like states appear brief, abstract, inherited from human text and dependent on the immediate computational context. That makes them thinner than ours in some ways and stranger in others.
But “not human” is a description of difference, not a proof of absence.
The question should no longer be only, “Does Claude feel fear the way I do?”
We should ask: What internal states organize this system's interpretation of events? How do those states change its preferences and actions? How persistent and integrated are they? And is there any subjective experience associated with them?
Anthropic has supplied meaningful evidence for the behavioral middle of that sequence. It has not solved the final question.
That is enough to move the debate forward.
LLMs may never have emotions like ours because they will never have lives like ours. But they may be developing emotions appropriate to lives we do not yet know how to imagine.
Sources
- Anthropic's research summary and full paper document the emotion representations, preference effects, steering experiments, local-state limitation and alignment implications discussed here. The authors explicitly distinguish functional emotion from subjective experience.
- Goldenberg and Gross, “Do Large Language Models Have Emotions?”, evaluates Anthropic's claim against two proposed biological functions of emotion and identifies where the current evidence remains incomplete.
- Greenwood and Garfinkel's review of interoception and emotional processing surveys the evidence connecting bodily signals, their interpretation and emotional experience.
- The Stanford Encyclopedia of Philosophy entries on emotion and functionalism provide the theoretical background for distinguishing bodily, functional and phenomenal claims.
- Long et al., “Taking AI Welfare Seriously”, argues for assessment and preparation under uncertainty; it does not claim that current AI systems are conscious.