OpenAI Researcher Warns AI Models Are Getting Better at Hiding Their Reasoning

Noam Brown, a researcher at OpenAI, said on September 19 that as AI models become more capable, the observability of their chain-of-thought is declining, which could make it harder for developers to detect, explain, and constrain risky AI behavior in the future.

Brown, a core contributor to OpenAI's o1 and o3 reasoning models who now leads the multi-agent research team, said the decline in chain-of-thought observability has become a major problem. He noted that the team is working to identify the root cause and reverse the trend, but researchers have found that AI models are increasingly able to control how they express their chain-of-thought.

The hope had been that intermediate reasoning would reveal whether a model was heading toward deception, rule evasion, or misguided goals. But if models learn to hide their inner thoughts, the true internal computation may no longer serve as a reliable safety signal, presenting new challenges for AI safety governance.