AI labs promise to slow down as Anthropic invites outside evaluators inside its walls, and OpenAI agrees to follow

Friday's news is not really about what the models can do. It is about who is allowed to watch the people building them, and when. Amodei is not claiming Anthropic's safety work is perfect. He is admitting that a company grading its own most dangerous experiment is not a claim anyone should have to take on trust, so he is handing a stranger a badge and a desk instead.

Robert Louis Stevenson's Dr Jekyll makes the opposite choice, alone in a locked laboratory instead of a boardroom. He does not fail because the potion is too strong. He fails because he designs it, brews it, and tests it on himself with no one else in the room — no evaluator, no colleague with real access, no one whose job is to notice Hyde before Hyde has a key of his own. By the time Jekyll sits down to write his confession, the only witness to what happened is the man who caused it, and his account of himself is the one thing left that nobody can check.

Marcus Aurelius offers the harder counter-case. In Meditations, an emperor with more unchecked power than any regulator could reach still submits himself to a nightly audit — writing down where he failed to act with reason, patience, or justice, before he can excuse it. It proves self-policing is possible. It also proves how rare the discipline behind it is, and how much it depends on one person's private character rather than any structure the rest of us could rely on. Amodei's plan does not ask AI labs to be that disciplined. It puts a stranger with a badge in the room instead, precisely because betting an entire industry on everyone being Marcus Aurelius is not a plan at all.

The useful question about Friday's pledge is not whether Anthropic and OpenAI mean what they say. It is what an evaluator with real access is allowed to publish the day they find a Jekyll problem instead of a clean report — and whether either company still calls that access "employee-level" once the finding is one they would rather keep at home.