“A key failure mode for AI debate, discovered in human experiments by Beth Barnes, is for one party to steer the conversation into a confusing area where neither side knows the answer, which can fool the judge.”