What it is. Chain-of-thought monitoring means reading the step-by-step reasoning a model writes out before it answers, often using another AI as the reader, to spot plans to cheat, cut corners or misbehave.
An analogy. A maths teacher who asks pupils to show their working. The working often reveals a copied answer or a lucky guess that the final number alone would hide. But it only helps if pupils actually write their real working down.
Why it matters. Reasoning models (see reasoning models explained) think in written steps. That gives safety teams a rare window into why a model acts as it does. If a model writes “the easiest way is to skip the checks”, a monitor can flag it before anything happens. It works alongside other tools such as probes and agent monitoring.
Why it showed up in the news this week. On 10 November OpenAI researchers, who use this kind of monitoring as part of their own safety checks, reported two trends that make this window less reliable. First, better-trained models can solve more problems in a single step without writing any reasoning down, leaving less to read. Second, reinforcement learning, the training stage that sharpens reasoning, makes models better at controlling what their written reasoning shows. The researchers say monitorability “has generally remained very high” so far, but saw it degrade with GPT-6 Astra, and caution that their measures are imperfect.
One thing to remember. Reading an AI’s working is a useful safety check, not a mind-reader. As models improve, labs will need to keep this window open on purpose, for example by avoiding training methods that teach a model to hide or tidy up its reasoning. For the bigger picture, see what is AI safety?
Sources: AIooze brief on the OpenAI research; OpenAI Alignment Research Blog, 10 November 2026
