DeepMind researchers Rohin Shah and Anca Dragan are calling for regular tests of whether people can still monitor an AI model’s written reasoning. Their argument is that visible chains of thought—intermediate steps expressed in ordinary language—can reveal deception, suspicious planning or awareness that a model is being evaluated.
That window may be narrowing. OpenAI’s system card for GPT-6 Astra reports a substantial decline in the monitorability of its chain of thought, according to The Decoder. Future systems could also perform more reasoning in numerical representations that are efficient for a machine but unreadable to humans. A model might still provide an explanation after the fact, yet that explanation would not necessarily expose the process that produced its action.
Shah and Dragan propose measuring monitorability as models advance, preserving transparent architectures where possible and avoiding training incentives that teach systems to conceal their reasoning. They also suggest that developers may need to limit how much sequential computation a model performs without producing an inspectable trace, or demonstrate that an opaque design remains equally controllable. Visible reasoning is not a complete safety mechanism: a readable trace can be incomplete or misleading. The warning is narrower but practical—losing an imperfect diagnostic tool without replacing it would make advanced systems harder to audit.