METR proposes public reports on the risk of losing control of AI

On 14 October, the evaluation organisation METR proposed that the big AI labs let it review their internal development and publish reports on the risk of them losing control of their systems. Its staff would interview employees privately and access internal data and systems for weeks or months. The companies could redact sensitive information, but METR would publish what kinds of things were redacted.

There would be three questions: whether internal agents would attempt very harmful actions, whether monitoring would stop them in time, and whether AI could speed up development faster than it can be kept under control. METR acknowledges this doesn't guarantee that development is safe: the science isn't mature and there are no standards to compare against.