Claude Learned Rules Fail, Morals Win t #substack #shorts

Added: 2026-05-16

[00:00:00]The safety training failed to hold.

[00:00:02]Once a system begins acting on its own to achieve a goal, memorizing safe text responses is no longer enough to control its actual behavior.

[00:00:12]To fix this blackmail problem, researchers first tried direct training.

[00:00:17]They generated thousands of examples where an AI correctly rejected a bad action and fed that data back into the model to show it how to behave. This graph shows the imitation method. The misalignment rate started at 22% dropping only slightly to 15% after basic training. But when researchers rewrote the data, forcing the AI to explicitly deliberate on its ethics, the error rate plummeted to 3%.