Over 40 researchers across OpenAI, Anthropic and DeepMind publish joint chain-of-thought monitorability paper

More than 40 researchers from OpenAI, Anthropic, Google DeepMind and other organisations, including the Center for AI Safety, co-authored “Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety.” The paper’s central claim was that reasoning models which “think” in natural language before answering give overseers a tool current architectures did not offer: a legible trace of the steps a model went through, which can in principle be read for signs of an intent to deceive, cheat or otherwise misbehave, even when the final output looks fine on its own.

The paper’s warning was that this property is contingent rather than guaranteed. Training pressures — reinforcement learning that rewards correct answers regardless of how they were reached, or optimisation that compresses reasoning into less human-readable forms — could erode the correspondence between what a model writes as its reasoning and what actually drives its output, without any single decision to abandon monitorability. The authors recommended that developers evaluate and publish monitorability scores for their models, treat monitorability as a factor in training and architecture decisions, and avoid changes that would degrade it without a compensating safety benefit.

A cross-lab paper with this many named co-authors from directly competing organisations was unusual; frontier labs typically publish safety research separately, if at all jointly branded. The paper drew public endorsement from figures including Ilya Sutskever, Geoffrey Hinton and John Schulman — none formal co-authors, but researchers whose support signalled the argument had support beyond the authoring labs’ current staff. It set a marker that later work, including automated monitoring benchmarks, referred back to when assessing whether specific models remained faithfully monitorable.