Abstract
A recent paper from Anthropic explores whether large language models (LLMs) possess emotions, concluding that Claude Sonnet 4.5 exhibits 'functional emotions' through internal representations of emotion concepts.
We evaluate this claim against what is known about emotions in biological systems, which serve two core functions:
- Context-sensitive interpretation: Emotions help individuals interpret complex situations in their environment.
- Reorganization of processing across multiple systems: Emotions prompt dynamic reorganization in response to these interpretations.
Anthropic's findings partially support the first function; however, the consistent and discrete emotional representations identified in Claude conflict with affective neuroscience findings, which indicate that human emotions are characterized by variable rather than uniform neural signatures.
For the second function, the evidence is mixed: Claude's representations can modulate output but fail to produce the dynamic reorganization of attention, decision speed, and motivational state that defines emotion in biological systems. We conclude by proposing what it would take for an LLM to truly possess emotions.
Blogger's Review: This article delves into the emotional representations of large language models. While there is some support for their emotional functionality, compared to the complexity of biological systems, LLMs still seem lacking in emotional depth. Future research should explore how to enhance emotional intelligence in these models.