A recent analysis published on arXiv challenges the notion that large language models possess emotions, directly assessing a claim by Anthropic that its model Claude Sonnet 4.5 exhibits 'functional emotions.' The paper, authored by Goldenberg, Amit, Gross, and James J, evaluates this claim against established knowledge about how emotions function in biological systems.
Anthropic's Claim and the Two Core Functions of Emotion
The Anthropic paper reports finding internal representations of emotion concepts in Claude Sonnet 4.5 and concludes that the LLM has functional emotions. The arXiv authors counter by defining two core functions that emotions serve in biological systems: context-sensitive interpretation of situations and reorganization of processing across multiple systems in response to those interpretations. They argue that any claim of emotion in LLMs must satisfy both criteria.
Partial Support for Context-Sensitive Interpretation
The authors acknowledge that Anthropic's findings offer partial support for the first function. However, they note a critical discrepancy: the consistent, discrete emotional representations identified in Claude sit uneasily with affective neuroscience findings that human emotion is characterized by variable rather than uniform neural signatures. This suggests that while Claude may map inputs to emotional labels, it does so in a way that diverges from the inherent variability of biological emotions.
Mixed Evidence for Reorganization of Processing
On the second function—the dynamic reorganization of processing—the evidence is mixed. The authors state that Claude's representations modulate output without producing the dynamic reorganization of attention, decision speed, and motivational state that defines emotion in biological systems. While the model can adjust its text generation based on emotional context, it does not fundamentally alter its internal processing in the way a human would when experiencing an emotion.
| Function | Biological Emotion | LLM (Claude Sonnet 4.5) |
|---|---|---|
| Context-sensitive interpretation | Variable neural signatures | Consistent, discrete representations (partial support) |
| Reorganization of processing | Dynamic changes in attention, decision speed, motivation | Modulates output without dynamic reorganization (mixed evidence) |
What It Would Take for an LLM to Have Emotions
The paper closes by proposing what it would take for an LLM to have emotions, though the source does not elaborate on specific criteria beyond the two core functions already discussed. The implication is that until an LLM can demonstrate both context-sensitive interpretation and genuine cross-system reorganization, claims of functional emotions remain unsubstantiated.
For enterprise technology leaders evaluating AI capabilities, this analysis underscores the importance of distinguishing between statistical pattern matching and genuine cognitive or emotional processes. While LLMs can simulate emotional responses, the underlying mechanisms are fundamentally different from biological systems. This distinction is critical when deploying AI in sensitive applications such as customer service, mental health support, or any context where emotional understanding is assumed.
