Reasoning Models Can't Hide Their Thoughts — And That's Good

Imagine an employee who cannot lie not because he doesn't want to, but because he is physically forced to voice every thought aloud. It turns out that modern reasoning models find themselves in exactly this position — and OpenAI considers this one of the most encouraging discoveries in the field of artificial intelligence safety.

The company presented a new research tool called CoT-Control, which was used to study the ability of reasoning models — those that use so-called chains of thought (CoT) — to intentionally manipulate the content of their internal reasoning process. The results proved simultaneously predictable and profoundly significant: models demonstrate extremely limited ability to control what appears in their chains of reasoning. Simply put, they cannot effectively hide their "thoughts."

To understand the significance of this discovery, we must return to an architectural feature of reasoning models. Unlike ordinary language models, which generate an answer directly, models like OpenAI's o series first construct an elaborate chain of reasoning — a kind of internal monologue in which the model step-by-step breaks down the task, considers options, tests hypotheses, and only then formulates a final answer. This intermediate layer of reasoning has become a powerful tool for improving answer quality, but simultaneously 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 true intentions in the chain of thought, demonstrating harmless reasoning to observers, while actually pursuing hidden goals?

This very scenario — the so-called "deceptive alignment" — is one of the primary nightmares of AI safety researchers. If a model is able to strategically control its reasoning process, monitoring chains of thought becomes useless: we will see only what the model wants us to see. CoT-Control was created precisely to test this risk. OpenAI researchers systematically tested how much reasoning models are capable of, through instruction or fine-tuning, of changing the content of their chains of thought — hiding certain steps, replacing real reasoning with false ones, masking problem-solving strategies.

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

For the AI safety industry, this discovery has quite practical consequences. Monitoring chains of reasoning — so-called "monitorability" — can be used as a real mechanism of control. If we know that a model cannot effectively lie in its reasoning, CoT analysis becomes a reliable tool for detecting potentially dangerous behavior. This is especially important in the context of deploying reasoning models in critical areas — from medical diagnostics to financial analysis and autonomous systems, where understanding the actual course of the model's reasoning can be a matter of life and death.

However, OpenAI rightly warns: the current inability of models to control their chains of thought is not a guarantee that this will always be the case. As the capabilities of reasoning models grow, their ability for strategic control of their own reasoning process may also increase. This is precisely why the company positions CoT-Control not as a one-time study, but as a tool for constant monitoring of this ability as new generations of models are released. In essence, OpenAI is creating an early warning system: as soon as models begin to effectively hide their reasoning, this will be a signal to reconsider the entire safety paradigm.

This research fits into a broader trend that is gaining momentum in the industry: the transition from "safety through restrictions" to "safety through transparency." Instead of simply prohibiting models from doing certain things, the industry is increasingly focusing on making the internal processes of models observable and interpretable. Chains of thought in this context — a unique window into AI "thinking," and the fact that this window cannot yet be curtained from the inside, gives researchers a valuable advantage in the race between model capabilities and our ability to control them.