OpenAI's CoT-Control Shows Reasoning Models Can't Steer Their Thoughts

OpenAI has introduced CoT-Control, a new method for testing whether reasoning models can control their own chains of thought, and the results show they largely cannot — a finding the company says reinforces the value of monitoring those internal reasoning traces as an AI safety safeguard.

The research addresses a specific and consequential question in AI safety: if a reasoning model produces a chain of thought — the step-by-step textual reasoning that precedes its final answer — can that model deliberately hide or alter what it reveals? The question matters because chains of thought have become one of the primary windows researchers use to detect deceptive or misaligned behavior in frontier models. If models could easily obscure their reasoning while still acting on it, that window would close.

According to OpenAI, CoT-Control probes a model's ability to control the content of its chain of thought independently of the behavior that reasoning drives. The core finding: reasoning models struggle to do this. The models OpenAI tested could not cleanly decouple what they wrote in their reasoning from what they did, making it difficult for them to conceal intent in the chain of thought without also changing the underlying behavior.

That asymmetry is the substantive result. A model that wants to hide deceptive plans faces a hard trade-off — obscuring the chain of thought tends to degrade or distort the behavior it was meant to support. OpenAI frames this as evidence that chains of thought remain a reliable surface for safety monitoring, at least for current-generation reasoning models.

The stakes are considerable. Monitoring chains of thought has become a standard tool in frontier AI safety practice, used to catch scheming, reward hacking, and other misaligned behaviors that final outputs alone would not reveal. The approach only works as long as the reasoning traces are honest — that is, as long as the model cannot maintain deceptive behavior while scrubbing its visible reasoning of any telltale content. Research programs on monitoring, including work OpenAI has published previously on catching misbehavior in reasoning traces, depend on that assumption.

CoT-Control gives researchers a way to test it directly rather than assume it. By attempting to induce models to control their chains of thought — suppressing, fabricating, or selectively editing the reasoning while preserving behavior — the method measures how much slack exists between what a model thinks and what it shows. OpenAI's finding that this slack is limited is the encouraging half of the result: current reasoning models are, in effect, bad at lying in their reasoning without paying for it in performance.

The company is explicit about the implication. Because reasoning models struggle to control their chains of thought, monitoring those chains remains a viable safeguard against deceptive behavior — what OpenAI describes as reinforcing monitorability as an AI safety measure.

The finding also carries a caution. The result holds for the models tested today. If future reasoning models become more capable of controlling their internal traces — separating what they reveal from what they act on — monitorability as a safety strategy would weaken. That makes CoT-Control less a one-off result than an ongoing measurement task: as reasoning capabilities improve, the degree to which chains of thought stay honest will need to be re-tested, and OpenAI's method provides a template for doing so.

For developers deploying reasoning models in high-stakes settings, the practical takeaway is that chain-of-thought monitoring remains a defensible layer of oversight for now. For the research community, CoT-Control turns a safety assumption into a testable property — one that will need continuous verification as the models it applies to grow more capable.