Embedded AI evaluators are a good start. But can they stop an unsafe run, independently verify the logs, and publish findings without company approval?
If they can only watch and report, we’ve added witnesses. We still need enforceable oversight.
This quoted post is unavailable.