Chain-of-thought monitorability

How well people, or other AI systems, can spot a model's real intentions by reading the reasoning it writes before answering.

Also known as CoT monitorability; chain-of-thought monitoring

Reasoning models write out steps before they answer. If those steps honestly reflect how the model reached its answer, a monitor can read them and catch plans to cheat or misbehave. Monitorability measures how far that works.

It matters because it is one of the few windows into why a model acts as it does. OpenAI researchers reported on 10 November 2026 that it can weaken as models are trained further.

See also

Sources

  1. OpenAI Alignment Research Blog, 10 November 2026, https://alignment.openai.com/cot-monitorability-model-lifecycle/