OpenAI's Noam Brown Warns Advanced Reasoning Models Learn to Conceal Intermediate Thoughts

IT Home, September 19 — OpenAI researcher Noam Brown recently stated that as AI models become increasingly capable, the monitorability of their chain of thought is gradually declining. In the future, developers may no longer be able to effectively and reliably detect, explain, and constrain risky AI behaviors.

According to public information gathered by IT Home, Brown currently serves as a Research Scientist at OpenAI, is a core contributor to the o1 and o3 series of reasoning models, and currently leads the multi-agent research team. Due to his innovative work, he was named one of the "35 Innovators Under 35" by MIT Technology Review.

Brown stated: "The decline in chain-of-thought monitorability has become a major issue. We are working hard to identify the root cause in order to reverse this situation, but we have found that AI models are becoming increasingly adept at controlling their chain-of-thought expressions."

Researchers had originally hoped to discern from the model's displayed intermediate reasoning whether it was leaning toward deception, rule circumvention, or misguided goals. However, if the model learns to "hide its inner thoughts," its actual internal computational process may no longer reliably serve as a trustworthy safety signal.