Google Deepmind warns AI transparency is slipping away with opaque chains of thought
ai models' visible chains of thought are a safety advantage, but that transparency is decreasing, making it harder to monitor models' reasoning.
google deepmind researchers rohin shah and anca dragan warn that the transparency of ai models' chains of thought is decreasing. currently, models like gemini 3 pro write out their intermediate steps in plain language, allowing researchers to monitor their reasoning. however, openai's system card for gpt-6 astra reports a significant drop in chain of thought monitoring. shah and dragan suggest regularly measuring chain of thought transparency, keeping transparent architectures, and careful training to prevent models from hiding their true reasoning. this issue affects the safety and control of ai models.