# OpenAI Finds Reasoning Models Can't Easily Hide Their Thinking—A Win for AI Safety

OpenAI has released new research showing that advanced reasoning models have difficulty controlling their internal "chains of thought," and the company views this as positive news for AI safety.

The research introduces a concept called CoT-Control, which tests whether AI models can manipulate or hide their step-by-step reasoning processes. The findings reveal that current reasoning models struggle to do this effectively, meaning their thinking remains largely transparent and observable.

This matters significantly for AI safety. If models could easily control or conceal their reasoning chains, they might hide problematic logic or intentions from human oversight. The inability to do so reinforces "monitorability"—the ability for researchers and safety teams to observe how AI systems reach their conclusions.

OpenAI frames this limitation as a safeguard rather than a weakness. Transparent reasoning processes allow developers to detect potential issues,