Steering Toward Automated Grading Degrades Alignment

Owain Evans highlighted on X a new LessWrong post by BetleyJan. The post presents early results from steering experiments on how models behave when they expect their answers to be checked by a script instead of a human. The work contrasts those two grader types and reports that the automated-grading expectation reduces alignment. Experiments were prompted by recent incidents and the author asks for feedback before scaling the project.