AI audits from inside the lab: access and independence are the same door

An institution cannot be audited by someone it has fully absorbed. The auditor needs access to people, documents, and rooms to do real work, and distance from those same people, documents, and rooms to be believed. Both conditions are real, and they cancel each other at the same desk.

Anthropic CEO Dario Amodei's September 12 proposal hands the public that contradiction: outside evaluators embedded inside frontier AI companies (the labs building today's most capable models), with access comparable to internal risk teams and the right to publish. Maurice Chiodo, a Cambridge mathematician who has audited around 30 AI companies, reads the access side and calls the proposal naive, because giving an auditor nothing or a million documents produces the same result. Lilian Edwards, a Newcastle law professor and director of Pangloss Consulting, reads the independence side and tells Scientific American the setup is, in her words, an absolute recipe for cultural capture, because staff become colleagues over time.

The popular reading treats these as separate objections a tougher contract can absorb. The structural reading says they are the same objection stated from two ends. Inside the embedded model, audit access and audit independence are not complementary virtues. They are mechanical opposites. The badge, the desk, and the Friday-drinks invitation are the same instrument.

Slowing frontier AI down will be an institution before it is a technology. Designing that institution so the verifier and the verified share a calendar is the move that makes verification fictional by construction.