AI disclosure.

What ArguFight's AI does

ArguFight uses large language models (LLMs) to judge structured debates between two users on a specified topic. Each AI judge reads both sides' written arguments and issues a verdict — a winner or tie — plus a short reasoning trail. Verdicts are routed across three independent model providers and combined via a documented algorithm.

What it isn't

AI verdicts are opinions on argument quality, not fact-finding. A verdict says which side argued the topic more effectively under our judging framework. It does not verify claims of fact, does not adjudicate truth, and should not be treated as authoritative on the underlying subject.

AI verdicts are not legal advice. Nothing an AI judge outputs constitutes legal counsel. If a debate touches a legal question, do not act on the verdict.

AI verdicts are not medical, financial, or professional advice. Same as above. If a debate touches a domain requiring a licensed professional, consult one.

AI can be wrong. Verdicts have documented error modes. Users can appeal any verdict; appeals are re-judged by a different model provider than the original.

Fact Check

Fact Check is a separate, optional tool — not part of judging. Debaters can run an in-debate Fact Check on a specific claim. It retrieves public sources (such as Wikipedia and web search) and an AI model returns a verdict — Accurate, Misleading, False, or Unverified — with links to the sources it used.

Fact Check is AI-generated and can be wrong. Verdicts are produced from retrieved sources by an AI model and may be mistaken, incomplete, or out of date. They are informational only, are not a substitute for independent verification, and should not be treated as authoritative. When the sources do not resolve a claim, the result is “Unverified” rather than a guess.

Fact Check never affects scoring. The AI judges do not see fact checks and do not reward or penalize factual accuracy — verdicts score argument quality, not who is factually correct. A fact check is a tool a debater may use in their own argument; it changes no scores or ratings.

Model providers in use

ArguFight routes verdicts across independent AI providers. The current provider set includes DeepSeek, Google Gemini, and Groq-hosted open models. Model versions are logged with every verdict; the version tag appears in the verdict card footer as Argufight v1.2 or similar.

When a model version is retired or replaced, prior verdicts remain attributed to the model that produced them. If a systemic bias is identified in a specific version, affected debates are re-adjudicated and the change is disclosed on the verdict independence page.

Claims we don't make

• We do not claim ArguFight's AI is unbiased. All LLMs carry the biases of their training data. We disclose this and use multi-provider routing to reduce the impact of any single model's biases on a verdict.
• We do not publish an accuracy percentage. Any specific "% accurate" claim would require substantiation on file per FTC guidance. We instead publish the model versions, provider mix, and appeals reversal rate on the trust page.
• We do not claim ArguFight replaces human judgment. It's a tool for competitive debate practice and entertainment, not a substitute for expert adjudication on any real-world question.

Appealing a verdict

Any user can appeal any verdict they participated in. Appeals are routed to a different model provider than the original judge. Appeal outcomes and reversal rates are tracked and disclosed on the trust page.

Reporting AI issues

If you believe a verdict shows a systemic issue (racist output, hallucinated fact, or product of prompt injection), email info@donkeyideas.com with the debate URL and specifics. Reports are logged in the admin audit log and investigated.