Anthropic emotion concepts paper offers rival read of LLM 'self-directed' pain claims

timfduffy · x · 2026-09-24

In the debate over the LLM pain-representation preprint, Tim Dufficy points to Anthropic's emotion concepts paper: the model tracks separate emotional representations for the current speaker vs another speaker, not for user vs assistant roles. The identified pain axis may thus reflect a 'current speaker pain' representation rather than genuinely model-directed, self-specific encoding — interesting, but not evidence for the stronger 'self-directed representation' claim.

Related event: Anthropic's Emotion Paper Sparks Debate on Self vs Other Representations(3 posts)→

Original post →

More from Research

Research channel →