DeepMind warns: decline in traceability makes AI control more difficult

Google DeepMind has issued a serious warning about the decrease in visibility of reasoning chains in artificial intelligence (AI). In a contribution from the recently established DeepMind Institute, researchers Rohin Shah and Anca Dragan emphasize that transparency is crucial for the control and safety of AI models. They argue that the ability to trace the thought processes of models is essential for timely identification of deception and problematic thought patterns.

One of the benefits of a visible "chain of thought" is that it enables researchers to understand not only the outcomes of a model but also the reasoning behind them. This clarifies the intentions and strategies of the models, which can help with timely interventions when critical situations arise. The researchers refer to the example of Gemini 3 Pro, where the visible thought steps provided important insights into the behavior of the model.

Nevertheless, Shah and Dragan note that transparency in AI is under pressure. The recently published system card for OpenAI’s GPT-6 Astra indicates a significant decrease in the ability to follow the model's thought processes. This lack of visibility can lead to difficulties in recognizing misleading information and problematic plans in a timely manner. The researchers point to the growing risks that arise when future models operate within opaque "number spaces" whose logic is not comprehensible to humans.

To counteract this development, Shah and Dragan advocate for a series of measures. They propose regularly evaluating the monitoring capacity of thought steps, maintaining transparent architectures, and being attentive during model training to prevent hidden reasoning strategies. These proposals aim to ensure the controllability of AI models, especially as the urgency of the situation becomes increasingly clear in light of the report on GPT-6 Astra. The contribution from DeepMind Institute highlights the need for a proactive approach in ensuring safety and oversight in the rapid development of AI technology.

Read the full article from AI Insider.