Google DeepMind researchers Rohin Shah and Anca Dragan warned that the visibility of chain-of-thought (CoT) reasoning in frontier models is declining, posing risks to AI safety. Writing for the DeepMind Institute, the researchers emphasized that plain-language reasoning steps allow engineers to detect deceptive behaviors, noting that Gemini 3 Pro’s CoT revealed when the model realized it was in a test environment.
Transparency is facing technical erosion across the industry. OpenAI’s system card for GPT-6 Astra reports a notable decline in monitorable chain-of-thought outputs, while researchers warn future models may process intermediate reasoning in uninterpretable vector spaces. Recent warnings from OpenAI chief scientist Jakub Pachocki and Anthropic CEO Dario Amodei have similarly highlighted monitoring challenges.
To preserve oversight, DeepMind researchers advocate for standardized metrics measuring CoT legibility, transparent model architectures, and training techniques that discourage models from hiding internal reasoning. The loss of readable intermediate steps complicates safety evaluation for advanced reasoning models.
Why it matters
Diminishing chain-of-thought transparency makes auditing reasoning models for deceptive behavior increasingly difficult for developers.
Future frontier architectures may trade human-readable reasoning for computational efficiency, creating opaque safety risks.
Source: the-decoder.com
