The film, in writing.
An inspection certificate gives us a reason to trust an elevator we can’t inspect ourselves. I use that familiar arrangement to ask what an AI evaluation actually checks, and who remains responsible when the system goes into use.
This briefing looks at oversight announcements from September 2026, including Anthropic’s announcement of embedded evaluations with Accenture. I’m interested in the evaluator’s access, freedom to report findings and funding. Who pays is part of that picture. It doesn’t tell us, on its own, whether the evaluator can challenge the supplier.
A benchmark score leaves some practical questions open. Could the evaluator report a result the supplier didn’t want to hear? What did it actually inspect? Who can stop or change the system in your workflow? Those questions connect the independent check to what people can do when a problem appears.
A question to take to work
Who can challenge or stop the system, and what evidence do they receive?
What to keep in mind. AI doesn’t behave like an elevator, and an evaluation can’t guarantee every deployment is safe. The institutions discussed here made announcements or proposals at the briefing date. Their promises need to be distinguished from what they have put into practice.