Reasoning models struggle to control their chains of thought, and that’s good

OpenAI has a new experiment on its hands, and the headline result is almost comforting in its bluntness: reasoning models are bad at controlling their own chains of thought. The company built what it calls CoT-Control, a setup designed to probe whether a model can deliberately steer the reasoning text it produces before landing on an answer. The finding is that it mostly can't, at least not with much precision.

That sounds like a shortcoming until you think about what it implies for safety work. Chain-of-thought monitoring has become one of the few practical windows researchers have into what a reasoning model is actually doing internally, rather than just what it outputs. If models were highly skilled at manipulating or disguising their own reasoning traces on command, that window would start to look a lot less trustworthy. A model that can rewrite its intermediate thoughts to say whatever it wants would make monitoring almost theatrical, useful for show but not for catching real problems.

OpenAI frames the struggle as evidence that today's reasoning chains still carry genuine signal about the model's process, not a performance staged for observers. That matters a great deal for the broader push toward interpretability, where the working assumption is that watching a model think out loud gives researchers an early-warning system for deception, reward hacking, or other behavior nobody wants to discover after deployment. If chains of thought were easily gamed, that early-warning system would be compromised right when it's needed most.

The research doesn't claim this protection is permanent. As models get more capable, and as they're trained under pressure that rewards hiding uncomfortable reasoning, there's no guarantee the same limitation holds. OpenAI's own framing treats this current inability as a fragile advantage worth studying now, while it still exists, rather than a settled fact about how reasoning models will always behave.