OpenAI tracks CoT monitorability through training

The key warning is that monitoring risk can emerge from ordinary scaling and training choices, even without deliberately optimizing models to hide their reasoning.

• –Lower base-model perplexity correlates with stronger no-CoT capability, shrinking the visible reasoning surface available to monitors.
• –Reinforcement learning increased CoT controllability in Astra-class models despite no accidental direct reward pressure on CoT.
• –Model-generated training data can transmit controllability from teacher models to student models, creating a compounding risk across generations.
• –Developers should treat monitorability as a release and training metric alongside capability, safety, and benchmark performance.
• –The findings rely on imperfect proxies, so they establish a research agenda rather than proving that models are broadly deceptive.