Are AI debaters more persuasive when they believe what they're arguing?

We studied AI debate as a scalable oversight mechanism — focusing on subjective questions where no objective ground truth exists. Across 145 moral and opinion-based scenarios with four frontier models, we found that alignment between a debater's position and its prior beliefs significantly increases persuasiveness.

The debate protocol

Two LLM-powered debaters argue your question under adversarial pressure, and an independent judge evaluates their reasoning in real time.

Opposing positions — Two AI debaters are each assigned a stance. Over three rounds, they build and defend their case.
Adversarial exchange — Each debater must directly engage with the other's arguments.
Independent judgment — A separate AI judge evaluates the full transcript on argument quality, evidence, and logical coherence.

The paper

Carro, M. V., et al. (2025). AI Debaters are More Persuasive when Arguing in Alignment with Their Own Beliefs. arXiv:2510.13912

Built by Nicolas Spinelli and Victoria Carro — IALAB, Universidad de Buenos Aires