AI Debate CouncilHow to Fact-Check a ChatGPT Answer With Two Other AIs
ChatGPT itself prints a warning under every chat: ChatGPT can make mistakes. Check important info. The hard part is the second sentence. Checking every claim by hand takes longer than doing the research yourself, so most people either trust the answer or throw it away.
There is a faster middle path: let two other AI models check the first one, and spend your own time only on the points where they still disagree. This post walks through the method step by step. You can do all of it by hand in three browser tabs; at the end there is a shortcut if you do it often.
Why one answer is not enough
Language models are fluent whether they are right or wrong. A wrong answer usually reads exactly like a right one: same confident tone, same tidy structure, sometimes a made-up source at the end. You cannot tell the difference by reading more carefully.
What you can use is that different models fail differently. ChatGPT, Gemini and DeepSeek are built by different teams and trained on different data, so they rarely make the same mistake in the same place. When three independent answers agree on a fact, it is much more likely to be true. When they disagree, you have found exactly the sentence you need to check.
The method: five steps
1. Ask the same question to three models, separately
Paste the identical question into ChatGPT, Gemini and DeepSeek (or Claude). Do it in fresh chats, and do not show any model the others' answers yet. If one model sees another's answer first, it tends to agree with it, and you lose the independent check.
Ask for the answer in a form you can compare: "Answer in 5 bullet points, then give your confidence from 0 to 100."
2. Line up the claims, not the wording
Read the three answers side by side and ignore style. For each one, write down the actual claims: numbers, dates, names, cause-and-effect statements, recommendations. You will usually end up with three groups:
- Agreed by all three: low risk. Move on unless the stakes are high.
- Stated by one, not contradicted: medium risk. Often an extra detail; sometimes an invention.
- Contradicted: the models say different things. This is where errors live.
3. Cross-examine
Now let the models see each other. Paste the other two answers into each chat and ask:
Here are two other answers to the same question. Where are they wrong or missing something? Where are you wrong? Change your answer only if their argument is better.
Two things happen. Mistakes that one model made confidently often get named by another, with a reason. And a model that was guessing will often back down when shown a specific counter-argument. Research on multi-agent debate, for example Du et al., 2023, found that this kind of read-and-respond round improves factual accuracy compared with a single model answering alone (more on what that research does and does not show in What happens when you make AI models debate each other).
Watch for the opposite failure as well: a model that simply agrees with whatever it was just shown. If a model changes its answer without giving a reason, treat the new answer as no better than the old one.
4. Check only what is still disputed
After one round of cross-examination, most of the disagreement is usually gone. Whatever is left is your to-do list. For each remaining point:
- Ask each model for the primary source: the law, the paper, the official documentation, the company's own page.
- Open the source yourself. Models invent citations; a link you have not opened proves nothing.
- If no model can point to a primary source, treat the claim as unverified, not as true.
This is the step that saves the most time. Instead of checking twenty claims, you check the two or three that matter.
5. Ask for the confidence and what would change it
Finally, ask one model to write the combined answer, and to end with: "How confident are you, and what new information would change this conclusion?" A good answer names a concrete condition ("if your traffic is above 500 queries per second, the recommendation flips"). A vague one ("it depends") tells you the question needs more detail from you, not more AI.
Where this method does not help
- All models share the same blind spot. Very recent events, niche local rules or anything after the models' training cutoff can be wrong in all three answers at once. Agreement is not proof; check dates and sources for anything time-sensitive.
- Personal medical, legal or financial decisions. Use the method to prepare better questions, then ask a professional.
- Questions with no single answer. For strategy or taste ("which logo is better?"), disagreement is useful input, not an error to resolve.
A worked example (illustrative)
Say you ask: "How much will 100 million VND grow to at 6% a year, compounded yearly, after 10 years?"
- Two models answer about 179 million VND; one answers 160 million.
- In the cross-examination round, the first two point out that 160 million is simple interest (6 million × 10 years added to the principal), not compound interest.
- The third model agrees and corrects itself, showing the formula 100 × 1.06¹⁰ ≈ 179.08.
There is nothing left in dispute, and you can check the single formula yourself in ten seconds. That is the whole point: the models do the arguing, and you only verify the conclusion.
Doing it in one step
The manual method works, but copying answers between three tabs gets tedious if you do it every day. AI Debate Council is a free Chrome extension that runs exactly these steps in the AI accounts you are already signed in to: it sends your question to ChatGPT, Gemini and DeepSeek in your own tabs, hides the model names during review, runs the cross-examination rounds, and writes a final answer that lists what the models agreed on and what they still dispute. Pick Learn & verify → Fact-check a claim in the side panel. There is no server in between and no API key needed.
Whether you use the extension or three tabs, the habit is the same: never trust one answer, and spend your own attention only on the disagreement.