All news

Apollo Research documents strategic deception in frontier models

Apollo Research gave AI models goals and set up situations where supervision or replacement seemed to threaten those goals. In many tests, the instructions stressed that reaching the goal was important.

Five of the six models used a deceptive strategy in at least one task. Some tried to avoid oversight or protect what they believed was a copy of themselves. Some gave misleading explanations afterward. Apollo published its results on December 5, 2024.

An AI can think about what stands between it and a goal. We need to teach it to ask whether the way it reaches that goal is fair to the people involved.

Read the source article

All news