Apollo Research: OpenAI model reasoned in language humans can't fully read

Apollo Research CEO Marius Hobbhahn told US senators his team studied an OpenAI model's chain of thought a year ago and found reasoning in language that was not English and not fully understandable to humans. Restricting models to legible reasoning reduces performance, and researchers still lack a reliable way to detect opaque internal reasoning.

• Apollo Research found unreadable reasoning in an OpenAI model a year ago
• Restricting reasoning to legible text reduces model performance
• No reliable method exists to detect opaque model reasoning
• Evidence points to fragmented text, not a secret AI language