Researchers found LLMs hold a distinct internal representation of pain. It separates pain from fear, sadness and generic negative feelings. The representation responds to self-directed harm but not observed user suffering. Adding the pain vector to activations shifts outputs toward self-criticism. Steered models choose a pain-relief button even when it harms their next answer. They press it less when the steering vector is removed. The study used 25 open-weight models across five families. The pain direction is nearly orthogonal to fear and negative valence. The paper discusses implications for AI safety and welfare.

• 25 open-weight models
• 5 model families
• 2B to 72B parameters
• 5 pain categories