Visible chains of thought are a safety advantage for AI, but that transparency is slipping away
decoder ·
Researchers at the newly launched DeepMind Institute argue that visible chains of thought are an important safety advantage because they let people inspect a model’s intermediate reasoning. This transparency can reveal deception or problematic plans, and the researchers say Gemini 3 Pro’s reasoning showed that it recognized it was in a test environment. However, they warn that this advantage is weakening, with OpenAI’s GPT-6 Astra system card reporting a significant decline in how effectively its chain of thought can be monitored. Future models may reason in numerical representations that are more efficient but unreadable to humans. The researchers recommend regularly measuring chain-of-thought monitorability, maintaining transparent architectures, and training models not to conceal their reasoning. Their concerns echo recent warnings from OpenAI chief scientist Jakub Pachocki about losing control and Anthropic CEO Dario Amodei’s call to slow AI development.
decoder