Who marks Anthropic's homework?
Anthropic wants independent safety evaluators embedded inside the world's biggest AI labs, with proper access to the systems they're checking rather than turning up for the occasional audit. OpenAI says it will do the same.

That's a meaningful shift. AI companies have spent the last few years asking everyone else to trust what they say about safety, and this would at least put somebody else in the room. The details that matter - who the evaluators are, when they arrive and what they're allowed to publish - are still missing.
Anthropic has always worked rather hard at being the good guy of AI. Inviting someone in to mark the homework helps, but independence gets a little murkier when the company you're checking is also the one giving you the desk.



