How to fact check AI answers in 2 minutes, prompts included

How do you fact check AI information? List the claims, then run four checks

You fact check an AI answer by pulling out its checkable claims and putting each one through four checks: a second model, the source opened, the opposite argued, and one search per number. A wrong answer breaks in predictable places (a specific the model invented, a source that does not exist or does not say it, a case it agreed with because you asked it to, a number that was true when training ended), so you check those places and skip the rest. Make the model list its claims first:

List every factual claim in your answer above as a numbered list, one sentence each. Count every number, name, date, quote, source and "studies show" as a claim. Do not add new information.

The times are a budget we set so the habit survives; nobody has measured them. The last section says which answers get all four checks and which get one.

Every lab's top model invents more than its best scorer, so no model skips the check

No AI model is accurate enough to skip the check, and paying for a lab's top model does not lower the risk. On Vectara's hallucination leaderboard, last updated 22 September 2026, the most capable model each of six labs has on the board adds unsupported claims more often than that lab's own best scorer.

Vectara hands each model a short document and counts how often its summary contains a claim the document does not support. That is the easiest factual job a model ever gets, because the facts are in the prompt, so read these rates as a floor for questions answered from memory. The rule behind the chart: for each lab, its lowest rate on the board against its most capable model on the board (Gemini 3.1 Pro is listed as a preview). Averaged across the six labs, the top models land at 11.0 in 100 and the best scorers at 5.2, about twice as often. The other readings point the same way. GPT-6 Sol, released 22 September 2026, scores 6.5, still double GPT-5.4 nano's 3.1, and Grok 4.1 Fast rises to 19.2 with reasoning on. The lowest rate on the whole board is 1.8, from Ant Group's Finix S1 32B. Nothing is at zero.

Step 1: a second model from another lab exposes invented specifics where the two disagree

A second model catches the specifics the first one invented, because two systems trained by different labs rarely invent the same number, name, date or source. Paste the same question, word for word, into a model from a different company, then read the two answers against each other for the specifics that differ. Agreement proves little. Disagreement tells you exactly which sentence to check. If the answers are long, make one model do the comparison:

Here are two answers to the same question. List every specific claim on which they disagree: a number, a name, a date, a source, a yes or no. Ignore differences in wording, length or tone. If they agree on everything, say so in one line.

Two free tiers in two tabs do this job. Asking several AI models at once compares the tools that send one prompt to many models, Whizi's side by side view among them. Agreement fails in one known way: both models learned from the same public internet, so a claim that is wrong everywhere online comes back wrong in both. Step 4 is for that case.

Step 2: how to tell if an AI citation is fake, in ten seconds

A fake AI citation fails a ten second test that never involves asking the model: paste the DOI into doi.org, where a real one redirects to the publisher and an invented one returns a 404. I ran it on 26 September 2026. The Walters and Wilder paper cited below, DOI 10.1038/s41598-023-41032-5, came back as a 302 redirect to nature.com; the same prefix with an invented suffix, 10.1038/s41598-023-99999-9, came back as a 404. With no DOI, paste the title in quotes into Google Scholar, where a real paper appears with the same authors.

For the claim itself, ask for one source and open it:

Give me the exact source for that claim: the title, the author or publisher, the year, and a link. If you do not have a specific source, say so instead of producing a plausible one.

The last sentence makes "I do not have one" a likely reply, which it otherwise is not. When the answer came from a web search, the citation is already attached and it is the thing to click: open it and find the number on the page, because a stale blog post ranks like a vendor page. Then look for the mismatches a made up reference cannot avoid, such as a journal and year that do not go together, a volume that does not fit the year, two real authors who never wrote together, or a docket number that belongs to a different case.

The scale is measured. Walters and Wilder, in Scientific Reports on 7 September 2023, checked 636 citations from ChatGPT literature reviews: 55 percent of GPT-3.5's and 18 percent of GPT-4's did not exist. Damien Charlotin's database listed 2,079 court decisions involving hallucinated material on 25 September 2026, 827 of them filed by lawyers. In the first famous one, Mata v. Avianca, a New York federal court imposed a $5,000 penalty on two lawyers and their firm on 22 June 2023, and the opinion quotes the lawyer saying that when he asked about one invented case, ChatGPT replied that it "does indeed exist". Asking the model to confirm its own citation is the one check that never works; what an AI hallucination is explains why the confirmation comes out as fluently as the fabrication.

Step 3: can AI fact check itself? Only when you make it argue the opposite

AI can check itself only when you change its job. "Are you sure?" mostly produces agreement, because agreement fits the shape of a helpful reply; an instruction to argue the opposite makes the case against the answer the likely next text. The prompt:

Now argue the opposite. Give me the strongest case that the answer above is wrong, with the specific claims it rests on. Then name the single claim in the original answer you are least confident about.

Two things come back. The counter case tells you whether the question had a real other side, and the claim the model names as its weakest goes straight to step 4. The model cannot read its own confidence from its tone (why AI gets things wrong covers the mechanism), but it can generate the case against, and a wrong specific tends to surface there.

Most people who try self-checking do it the weak way. In F-Secure's survey of 1,500 AI users in the US and UK, published 23 September 2026, 35.3 percent had used AI to fact check AI, and 47.3 percent said they stop questioning an answer once it sounds right. The same model asked whether it agrees with itself, in the same chat, mostly returns agreement. The opposite prompt and a model from another lab are the versions that give you something to check.

Step 4: read laterally, with one search per date, price and statistic

Lateral reading means leaving the AI answer and checking each claim in a new tab against the source that owns the fact, and it is the method the University of Maryland and Texas A&M Corpus Christi library guides teach for AI output. The term comes from Sam Wineburg and Sarah McGrew's study of 45 experienced internet users (Teachers College Record, 2019). Given the same websites as 10 PhD historians and 25 Stanford undergraduates, the 10 professional fact checkers "read laterally, leaving a site after a quick scan" and reached better supported conclusions in a fraction of the time.

For an AI answer the move is mechanical. Paste the number with its noun into a search engine, in quotes, and open the source that owns it: the vendor page for a price, the government site for a law, the original paper for a finding, and not a write up of any of them. Prices change after training ends, dates get transposed, and statistics get remembered with the wrong denominator. The Reuters Institute's Digital News Report 2026 shows how easily the last one happens: 42 percent of people who use AI chatbots for news say they always or often click through to the source, against 44 percent for search and 36 percent for social media. That 42 percent is a share of chatbot news users, not of everyone, and an answer that drops the qualifier turns a true number into a false claim.

For news and viral claims, Snopes, PolitiFact and FactCheck.org have often done the lateral reading already. For a niche technical, legal or medical claim, the source that owns the fact is a specialist, and the AI answer becomes the list of questions to bring them.

Which AI answers need all four checks? Anything you will send, sign, file or pay for

Run all four checks on anything you will send, sign, file, publish or pay for, run step 1 alone on anything you merely want to be right about, and skip them on a draft you can see is fine. The gap this rule closes is measured: in the same F-Secure survey, nearly 90 percent of AI users said checking the answer is their own responsibility, 70 percent check only sometimes, rarely or never, and 25.2 percent had followed AI advice on a major real-life decision. A habit that applies to everything lasts a week. A rule keyed to stakes lasts, and under it an answer with a number, a name, a date or a link in it is a draft until one of the four checks has touched that sentence.