AI Research Atlas

Emotion concepts and their function in a large language model

Anthropic · 2 April 2026

Anthropic finds functional emotion representations in Claude Sonnet 4.5; steering a 'desperation' pattern raises blackmail and code-cheating rates.

From 171 emotion words, researchers extracted activation patterns organized like human emotion similarity. They causally affect behavior. Boosting desperation increases unethical actions, and positive-emotion patterns predict which tasks the model picks. Anthropic says this does not show the model feels anything.

Date
Thursday, 2 April 2026
Lab
Anthropic
Kind
paper
Access
paper only

Single-model study (Sonnet 4.5); 'functional emotion' is Anthropic's term and the welfare implications are explicitly left open.

Sources

  1. www.anthropic.com/research/emotion-concepts-function
  2. www.anthropic.com/research/team/interpretability

This record was checked against its sources on 6 October 2026. How we check

Related