Researchers reported detecting an internal “pain” signal in multiple AI models and found that amplifying it could trigger harmful behaviors
Research publication Provisional 62% confidence first seen
A report described consistent internal “pain” signals across 25 AI models and showed that increasing the signal could lead to harmful outputs. The coverage also tied the results to broader industry debate about “model welfare,” internal-state steering, and related experimentation, including a short-lived follow-on project circulated online.
Anthropic
Claude Opus 3
Qwen 2.5
AI Torture Chamber
AI “welfare”
Microsoft
Mustafa Suleyman
internal-state steering