A test may be of limited use anyway. The AI might be deceiving its creators (false negative). Or someone might train an AI with the goal to trigger a false posi

A test may be of limited use anyway. The AI might be deceiving its creators (false negative). Or someone might train an AI with the goal to trigger a false positive. It would only be useful if you had a cooperative AI not designed to game the test. In which case you either are still very far below the tester's capabilities in many relevant domains or you have solved the alignment problem.