OpenAI research shows CoT monitorability decreases as model intelligence scales
November 11, 2026
OpenAI findings indicate that models become harder to monitor via Chain of Thought as they scale toward higher intelligence. The study specifically examines the relationship between model training lifecycles, no-CoT capability, and CoT controllability.
HOW THIS AFFECTS YOU
●
researcherYou need to develop new ways to check model reasoning as scaling makes current CoT methods less reliable.
●
policyIt might get harder to audit how models make decisions as they get smarter.