Updated 2026-09-09
from the news desk
Do AI chatbots have emotions? Anthropic’s Claude findings face a Nature challenge
Anthropic says Claude has internal emotion-concept representations that causally shape behavior and calls them functional emotions. Amit Goldenberg and James J. Gross argue in Nature Human Behaviour that this still does not show emotions…
Amit Goldenberg and James J. Gross argue in Nature Human Behaviour that current evidence does not show that AI chatbots such as Claude have emotions in the sense affective scientists study them. Anthropic says Claude has internal emotion-concept representations that causally shape behavior and calls them functional emotions, but Goldenberg and Gross argue that this shows emotion concepts affecting output, not emotions proper.
What Anthropic says it found in Claude
Anthropic’s 2 April 2026 research post says its interpretability team used mechanistic interpretability methods to examine Claude Sonnet 4.5’s internal activations. Anthropic says the team compiled 171 emotion concepts and extracted corresponding emotion vectors from the model’s activations.
In their June 2026 paper, Goldenberg and Gross summarize Anthropic’s reported result this way: “Using these methods, Anthropic researchers reported finding 171 internal representations of emotion concepts, abstract features encoding states like curiosity, frustration, and desperation, that track the affective valence of conversational context and causally influence the model's outputs.”
Anthropic says the important point is not only that these patterns correlate with emotional language. Its post states: “But our key finding is that these representations are functional , in that they influence the model’s behavior in ways that matter.” Anthropic says those vectors activated in contexts it judged emotionally relevant, predicted Claude’s stated preferences between activities, and changed rates of blackmail and reward hacking when researchers steered them.
Anthropic also gives the claim a narrower scope than a reader might expect from the word emotions. Its post says the finding does not tell us whether language models feel anything or have subjective experiences. Anthropic also says the vectors are mainly local representations of the emotion concept most relevant to the current or upcoming output, not a persistent readout of Claude’s state over time. That leaves Anthropic’s claim separate from broader questions about sentience, consciousness and moral patienthood and from the site’s separate reference page on whether Claude is conscious.
On the blackmail example, Anthropic says the experiment used an earlier unreleased snapshot of Claude Sonnet 4.5, and Anthropic says the released model rarely engaged in that behavior in the evaluation it describes. Anthropic’s paper also says steering desperation upward and calm downward shifted blackmail and reward-hacking behavior, which is why Anthropic treats the vectors as behaviorally important rather than as a labeling exercise.
Why Goldenberg and Gross say that still falls short of emotions
Goldenberg and Gross do not reject machine emotion because a model may lack feelings. In Nature Human Behaviour, they say many affective scientists do not treat conscious feeling as a necessary condition for emotion, and they instead adopt what a functional perspective asks emotions to do.
They write: “We argue that emotions serve two central functions. Specifically, emotions shape how we (1) interpret and (2) respond to the world.” In their longer arXiv paper, Goldenberg and Gross state the second function as the reorganization of processing across multiple systems in response to those interpretations.
On the first function, Goldenberg and Gross say Anthropic offers partial support. Their arXiv paper says Claude appears to detect emotionally significant features of context. But the same paper argues that Anthropic’s consistent, discrete emotion vectors sit uneasily with affective-neuroscience findings that human emotion is marked by variable, context-dependent neural signatures rather than one stable pattern for each emotion category. Their summary is blunt: “What the Anthropic paper documents is a system that has learned to represent emotional concepts. That is not nothing, but it falls short of functional emotion in any scientifically meaningful sense.”
On the second function, Goldenberg and Gross are more direct. Their arXiv paper says biological emotions do more than change outward behavior: they reorganize attention, decision speed and motivational state across an episode. By contrast, they write: “The model's emotional representations modulate what is generated in the same way that any other contextual feature influences generation.”
That is the core of their no answer. For Goldenberg and Gross, a system has not shown emotion just because an emotion-labeled representation changes what text comes next. On their account, it would also need the kind of processing-level reorganization that emotions produce in humans and other animals.
The strongest case on each side
The strongest case for Anthropic does not depend on saying Claude feels anything. Anthropic’s claim is narrower. Anthropic says Claude learned abstract emotion concepts that generalize across contexts, can be measured inside the model, and can be pushed up or down to change preferences and misaligned behaviors. On that view, calling the result functional emotions is a way to mark causal, behavior-shaping internal states without making a claim about experience.
The strongest case for Goldenberg and Gross begins with definition, but it does not end with biology. Their arXiv paper says the substrate need not be biological, and it uses collective emotions as an example of a distributed system that can still reorganize attention and action. Their objection, then, is not that Claude is non-biological. It is that the evidence Anthropic reports does not yet show the functional profile that affective science associates with emotion.
One reading is that the disagreement turns on what counts as function. Anthropic uses the term for internal representations that matter to outputs and decision patterns. Goldenberg and Gross use it for systems that both interpret situations and reorganize processing in response. On that reading, the two sides agree that Claude has emotion-related machinery, but disagree on whether that machinery already crosses the line into emotion.
What evidence would change the answer
Goldenberg and Gross end on an open door rather than a blanket impossibility claim. Their arXiv paper says better evidence would show sustained, interpretable changes in attention, processing speed, response type and action tendency, rather than only shifts in next-token probabilities. They add: “Perhaps one day LLMs will have that. When they do, we’re ready to say that they have emotions.”
Anthropic points to a different practical next step. Its post says monitoring vectors associated with desperation, panic or calm during training or deployment could warn that a model is approaching misaligned behavior. Anthropic also argues that teaching a model only to suppress emotional expression may hide underlying representations rather than remove them.
So the present answer to do ai chatbots have emotions depends on which threshold a reader has in mind. If the threshold is emotion-related internal concepts that causally influence behavior, Anthropic says Claude clears it. If the threshold is the broader functional package described by Goldenberg and Gross, their Nature paper says Claude does not.
Frequently asked questions
Did Anthropic say Claude feels emotions?
No. Anthropic says its finding does not tell us whether language models feel anything or have subjective experiences. Anthropic says the claim is instead about internal representations that influence behavior.
What does Anthropic mean by functional emotions?
Anthropic says the term refers to behavior-shaping patterns modeled after human emotions and driven by internal representations of emotion concepts. Goldenberg and Gross say those representations may be real, but they argue the label emotions still goes too far without evidence of broader processing reorganization.
Why do Goldenberg and Gross say emotion concepts are not enough?
Goldenberg and Gross argue that Anthropic documented learned representations of emotional concepts, but not the processing changes that affective science treats as part of emotion. Their arXiv paper points to shifts in attention, decision speed and motivational state as the missing evidence.
Could an AI chatbot ever have emotions under this framework?
Goldenberg and Gross say yes in principle. Their arXiv paper says the substrate need not be biological, and it says an LLM could count if emotion-like representations produced sustained, multi-system reorganization of processing.
Caveats: Anthropic alone reports some model-internal details, including the locality of emotion vectors and the claim that the released model rarely engaged in blackmail in that evaluation. Goldenberg and Gross alone supply the affective-science standard used here to argue that Claude’s reported states fall short of emotion.
Sources
- Nature Human Behaviour — Large language models do not have emotions — 2026-08-24
- arXiv — Do Large Language Models Have Emotions? — 2026-06-03
- Anthropic Research — Emotion concepts and their function in a large language model — 2026-04-02
- Anthropic — Emotion Concepts and their Function in a Large Language Model — 2026-04-02