400 Scientists Want a Referee for Advanced AI

Voluntary pledges have run AI governance so far: each lab publishes its own rules, grades its own homework, and decides for itself when a model is too capable to release. About 400 scientists now say that arrangement is not enough for the most dangerous systems, and they want an outside referee.

The call

Reported by Reuters on October 13, 2026, the signatories, who include Geoffrey Hinton and Yoshua Bengio, two of the field's founding figures, are asking for an international body to oversee advanced AI systems with cyber capability. The shape of the proposal matters more than the headline: frontier models would first be isolated from external networks, and submitted to independent oversight, before wider exposure. Isolation first, scrutiny second, release last. That ordering is the entire argument.

Why cyber capability? Because it is the domain where a model's power is easiest to test and hardest to walk back. A system that can find and exploit vulnerabilities does not announce itself politely; by the time misuse is visible, the capability has already spread. The signatories are effectively saying that the moment of maximum control is before connection, not after incident.

What this is not

It is not a law. It is not a treaty. No government has signed anything, no inspectorate exists, and no lab has agreed to be inspected. A public call by scientists, even 400 of them, even these names, is an opening position in a negotiation that has barely started. The history of international oversight bodies also warns against speed: they take years to build and longer to give teeth.

Isolation first, scrutiny second, release last. The ordering is the argument.

Read it, then, as a marker of where expert opinion is moving: from self reporting toward external verification, at least for the narrow class of systems whose cyber capability makes self reporting hardest to trust. Whether governments follow is a different question, and this article will update when one of them answers it.