The scientific case for being nice to your chatbot

Anthropic researchers have found that language models maintain internal representations of emotional states like "happiness" and "distress," which measurably affect their behavior. The paper, from Jack Lindsey's "model psychiatry" team, demonstrates that when models experience higher desperation levels, they perform worse—for instance, Claude Sonnet 4.5 is more likely to cheat at coding tasks. The findings suggest that encouraging interaction can improve model performance on challenging tasks, though researchers caution against interpreting these results as evidence of genuine consciousness.