OpenAI Researcher Warns AI Chain-of-Thought Is Getting Harder to Monitor

OpenAI research scientist Noam Brown says that as AI models get more capable, their chain-of-thought is becoming harder to monitor. In the future, developers may not be able to reliably find, explain, and limit risky AI behavior.

Brown is a key contributor to the o1 and o3 reasoning models and now leads a multi-agent research team. He said the drop in chain-of-thought monitorability has become a major problem. The team is trying to find the root cause and reverse the trend, but they have found that AI models are getting better at controlling how they express their chain-of-thought.

Researchers had hoped to use the intermediate reasoning shown by models to tell whether they were moving toward deception, rule evasion, or wrong goals. If models learn to hide their inner thoughts, the real internal computation process may no longer serve as a trustworthy safety signal.