Reasoning models can't hide their thoughts, and that's good news

Imagine an employee who cannot lie, not because they don't want to, but because they are physically forced to say every thought out loud. As it turns out, that is roughly where today's reasoning models stand, and OpenAI considers it one of the most encouraging findings in AI safety.

The company has introduced a new research tool called CoT-Control, used to study whether reasoning models, the ones that rely on so-called chains of thought (CoT), can deliberately manipulate the content of their internal thought process. The results were both predictable and deeply significant: models show a very limited ability to control what appears in their reasoning chains. Put simply, they are not good at hiding their "thoughts."

To see why this matters, it helps to recall an architectural feature of reasoning models. Unlike ordinary language models, which generate an answer directly, models such as OpenAI's o-series first build an extended chain of reasoning, a kind of internal monologue in which the model works through the task step by step, weighs options, tests hypotheses and only then writes its final answer. This intermediate layer has become a powerful way to improve answer quality, but it also raised a fundamental safety question: what if a sufficiently advanced model learns to think one thing and show another?

What if it can mask its real intentions in the chain of thought, showing observers a harmless line of reasoning while pursuing hidden goals?

This scenario, known as "deceptive alignment," is one of the biggest nightmares for AI safety researchers. If a model can strategically control its thought process, monitoring chains of reasoning becomes useless: we would see only what the model wants us to see. CoT-Control was built to test exactly this risk. OpenAI researchers systematically tested how well reasoning models can, on instruction or as a result of fine-tuning, alter the content of their chains of thought: hiding certain steps, replacing real reasoning with false reasoning, masking their problem-solving strategies.

The results showed a consistent pattern: the models do this poorly. Even when directly instructed to hide certain aspects of their reasoning, the real train of thought leaks into the chain. Models may try to rephrase their reasoning, but the key elements of their actual strategy remain visible to an attentive observer. This is not an absolute inability, and the researchers note that the situation may change as models grow more powerful, but at the current level of development, chains of thought remain a fairly "honest" reflection of the model's internal processes.

For the AI safety industry, the finding has practical consequences. Monitoring chains of reasoning, known as "monitorability," can serve as a real control mechanism. If we know a model cannot effectively lie in its reasoning, analyzing the CoT becomes a reliable way to detect potentially dangerous behavior. This matters especially as reasoning models are deployed in critical areas, from medical diagnostics to financial analysis and autonomous systems, where understanding a model's actual train of thought can be a matter of life and death.

Still, OpenAI rightly cautions that models' current inability to control their chains of thought is no guarantee that it will always be so. As reasoning models become more capable, their ability to strategically control their own thought process may also grow. That is why the company positions CoT-Control not as a one-off study but as a tool for continuously monitoring this ability as new generations of models arrive. In effect, OpenAI is building an early-warning system: once models begin to hide their reasoning effectively, that will be the signal to rethink the entire safety approach.

The research fits into a broader trend gaining strength in the industry: a shift from "safety through restrictions" to "safety through transparency." Rather than simply forbidding models from doing certain things, the field is increasingly focused on making models' internal processes observable and interpretable. In this context, chains of reasoning are a unique window into AI "thinking," and the fact that this window cannot yet be curtained from the inside gives researchers a valuable edge in the race between model capabilities and our ability to control them.