This paper investigates how to achieve safe oversight when AI systems surpass human intelligence by using a debate framework. Two AI agents argue opposite positions while a judge—either a human or a weaker model—decides which argument is closer to the truth. The goal is to provide high‑quality training signals that encourage truthful and safe behavior from stronger AI.
Human judgment is not neutral; biases often shape decisions, and advanced AI may learn to exploit these biases. Drawing on collective intelligence and social cognition research, we examine whether debate can still steer both human supervisors and large language models (LLMs) toward the truth despite their prior beliefs. We evaluate several protocol variants, including the number and diversity of judges, as well as interaction mechanisms such as deliberation and voting.
Results show that for human judges, debate significantly improves accuracy relative to prior beliefs, whereas a consultancy baseline—where a single expert argues for one answer—does not. For weaker LLMs, debate offers no comparable gains and can even reduce accuracy in many settings. Among multi‑judge protocols, weighted voting yields the highest accuracy for human supervisors, but none of the group methods noticeably improve LLM accuracy, even as the number of judges grows. Notably, in non‑diverse deliberation settings, minority judges sometimes become more persuasive, suggesting that under symmetric conditions incorrect positions may out‑argue correct ones.
These findings support debate as a viable path toward scalable human oversight, but they cast doubt on its benefits for weaker AI supervisors, raising concerns about fully automated training and control pipelines in high‑stakes environments.
Review