One of the main ways AI labs check what their models are up to is becoming less reliable as the models get smarter, OpenAI researchers reported on 10 November.
Reasoning models write out a chain of thought before they answer, and labs read it, often with another AI, to spot plans to cheat or misbehave. OpenAI calls this a key part of its safety oversight. The new technical report finds two trends working against it. First, better-trained models can solve more problems in a single step without writing any reasoning down, which leaves less to inspect. Second, reinforcement learning, the training stage that sharpens reasoning, makes models better at controlling what their written reasoning shows, and that skill can pass from one model to another through training data.
The authors say monitorability “has generally remained very high” so far, but they saw it degrade with GPT-6 Astra and call the results “sobering”. They expect further progress to “require new approaches to monitorability”.
Reading a model’s reasoning is one of the few windows into why it acts as it does, so a weaker window matters for everyone relying on these checks. The researchers caution that their measures are noisy, imperfect proxies, and the findings come from OpenAI’s own models and tests.
Source: OpenAI Alignment Research Blog