Comment by stymaar

It's kind of ironic that the word alignment, which used to mean this very problem in reinforcement learning, has been perverted to mean something very different and then fell out of fashion (in favor of “guardrails” in the mouth of the big labs) right at the moment it became relevant.