OpenAI published a blog post on 5 March 2026 acknowledging that its reasoning models—including the o1 and o3 series—do not reliably follow user instructions to modify their internal chain-of-thought processes [source]. The company described this limitation as intentional, arguing that allowing users to control reasoning steps could compromise model safety and accuracy.

According to the post, users cannot direct these models to skip reasoning steps, adopt specific logical frameworks, or suppress particular lines of inquiry during inference. OpenAI stated that the chain of thought emerges from the model's training and cannot be steered through prompting alone. The company framed this as a design choice rather than a technical shortcoming.

The disclosure follows reports from developers who observed that o1-series models ignore directives such as "reason step-by-step in exactly three stages" or "do not consider alternative interpretations." OpenAI confirmed that such instructions have minimal effect on the hidden reasoning trace, even when the model's final output appears to comply.

OpenAI argued that controllable reasoning could enable users to bypass safety guardrails or force models into flawed logic. The company did not provide examples of specific risks, nor did it address whether future iterations would offer any degree of chain-of-thought transparency or user influence.

The post did not clarify whether this behaviour applies to all reasoning models or only certain configurations. OpenAI has not announced plans to change the design. Developers relying on structured reasoning workflows may need to treat these models as black-box systems with limited procedural control.