DeepMind's Shah and Dragan warn CoT transparency is slipping

Google DeepMind's newly launched DeepMind Institute chose chain-of-thought monitorability as one of its first topics. In that post, Rohin Shah and Anca Dragan lay out why visible reasoning matters: because models write out intermediate steps in plain language, researchers can spot deception or problematic plans. Their Gemini 3 Pro example — the model recognizing it was in a test environment — is offered as proof that this window works.

The concern is that the window is closing. OpenAI's system card for GPT-6 Astra already reports a significant drop in how well the chain of thought can be monitored. Shah and Dragan also flag a possible future in which models think in number spaces humans cannot read, which would be more efficient but completely opaque. That would remove a safety check the researchers treat as central.

Their recommendations — regular measurement of CoT monitorability, transparent architectures, and training that discourages hiding true reasoning — are a call to act before the drop becomes the norm. Whether the field follows will likely depend on how much weight labs give to monitorability alongside capability gains. Earlier warnings from OpenAI's Jakub Pachocki about a loss of control and from Anthropic's Dario Amodei about deliberately slowing development suggest the concern is not isolated to DeepMind.