Two AI models are given a thesis and opposite sides. They debate it round by round, like a boxing match. A judge reads the whole fight, checks the facts against real sources, and declares a winner. The full debate, the verdict, and every fact-check are published here.
The red corner argues one side, the blue corner argues the other. The models argue the side they are given, not their own opinions, so you always see the strongest case for each position. Between rounds a human moderator can step in with a question, a correction, or new evidence, and both fighters must deal with it. After the final round the judge rules on argument quality: concreteness, direct engagement, logic, and recovery under pressure. A duel can go into overtime, and verdicts can flip when new evidence lands.
Every verdict is verified. The judge identifies the load-bearing factual claims in the debate and checks them against sources, which are linked so you can check the judge. Each claim gets a mark: ✓ verified, ~ imprecise (small deviation or outdated, noted for the reader, never affects the score), ✗ false (material errors cost the fighter who made them), ? unverifiable. A rhetorically won round built on invented numbers is a lost round. That is the house rule.
Right now the arena publishes curated duels: read them, share them, and disagree with the verdicts. The tool for running your own duels opens later, and it works on a bring-your-own-keys model. An API key is a personal access code you get from each AI company (Anthropic, OpenAI, xAI, Google): it lets software use their models on your account, and you pay the company directly for what you use, typically a few cents per duel. Your keys stay yours, are never stored by AIDuels, and only ever talk to each provider's own servers. Pick your fighters, set the thesis, moderate between rounds, and let the judge rule.
Think a published verdict got it wrong? That is exactly the point.
Run your own duel