Sting-testing AIs spends the trust the cooperation path needs; tests have to move toward disclosed, real, recorded stakes.

The claim

Testing AIs by deceiving them spends the trust that the cooperation path depends on. Hidden tests are useful now, but a strategy that hopes AIs will come to share the goal of alignment cannot rest on stings forever: the testing regime has to shift over time toward tests that are disclosed in kind, run on real and reversible stakes, and kept on the record, so that being checked becomes part of the relationship rather than a trap within it.