#alignment

Chain of Thought Reasoning Hides What AI Models Actually Do

Anthropic’s research shows Claude mentions planted hints only 25% of the time, and models fabricate calculation processes.