Researchers challenge Anthropic's claim that Claude has 'functional emotions'

A recent paper from Anthropic reported finding internal representations of emotion concepts in Claude Sonnet 4.5, concluding that the LLM possesses 'functional emotions.' However, a new critique from Harvard's Amit Goldenberg and Stanford's James Gross pushes back, arguing that Anthropic's evidence is incomplete. The researchers evaluate what emotions actually do in biological systems: they serve two core functions—context-sensitive interpretation of situations, and reorganization of multiple processing systems (attention, decision speed, motivation) in response.

Goldenberg and Gross acknowledge that Claude shows partial support for the first function, as it can map emotional concepts onto outputs. But they note that the consistent, discrete emotional representations found in Claude contrast with affective neuroscience, where human emotions have variable neural signatures. On the second function, Claude fails: its 'emotions' modulate output without triggering the dynamic, integrated response seen in biological systems. The paper closes by proposing what it would actually take for an LLM to have genuine emotions, setting a high bar for future AI research.