The judges were given a question about a science fiction story and two possible answers. One answer was right. The other was the most tempting wrong answer the question writers could come up with. The judges were not allowed to read the story.
The stories came from Project Gutenberg, mostly from the 1950s to the 1970s, so it was unlikely anyone had read them before. The questions were drawn from a set chosen to be hard, ones that people who read the whole story without a time limit had all answered correctly and called unambiguous.
The experiment, run by researchers from University College London, Anthropic and other groups and published in 2024, was a stand-in for a problem AI developers expect to face: people having to check the work of systems that know more than they do. Here the gap was simply information. The AI models acting as experts, among them GPT-4-Turbo, had the full story. The human judges, 21 trained crowdworkers paid $35 an hour, had only what the experts told them.
The experts had one constraint. They could quote the story, and a tool checked each quote against the text and marked it verified or not. Judges were told to trust only verified quotes.
With no expert at all, judges answering from the question alone got 60% right on a choice between two answers.
In one setup, called consultancy, a single expert argued for one of the two answers, which might be the wrong one, and the judge could ask it questions. Judges got 78% right. Their most common mistake was failing to get the key quotes out of an expert arguing for the wrong answer.
In the other, called debate, two experts argued for opposite answers over three rounds. Judges got 88% right.
The researchers also tried making the experts more persuasive. In debate, more persuasive debaters led judges to the truth more often. With a single consultant, more persuasive experts made judges less accurate.
AI judges, standing in for the human ones, showed the same pattern at lower levels: 48% with no expert, 54% with one, 76% with debate.
The setup was artificial, and the authors noted that having more information is only one way future AI might outstrip its overseers. But across more than a thousand hours of human judging, two experts made to argue, under rules that let the judge check the evidence, did better than listening to one expert arguing for an assigned answer.