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

OpenAI has introduced a method called CoT-Control to test whether reasoning models can alter or suppress their chain-of-thought (CoT) outputs. Their experiments revealed that these models struggle to control their internal reasoning when prompted to do so—even explicit instructions fail to fully suppress reasoning traces. According to OpenAI Blog, this inability to "hide" thoughts is actually good news for AI safety: it ensures that CoT reasoning remains monitorable and auditable, providing a built-in transparency safeguard. The practical implication for developers building AI workflows is that they can rely on the visibility of a model's reasoning steps for debugging and verification, especially in high-stakes applications. However, builders should also note that attempts to force models to skip reasoning may lead to degraded performance, so workflow designs should embrace rather than circumvent this transparency.