How to fact-check an AI answer
Short answer
To fact-check AI, check five things in order: the claims the answer rests on, the sources it cites, its numbers, its dates and its quotes. Each takes a minute with a free tool. It matters: in the EBU/BBC study, 45% of AI news answers had at least one significant issue (October 2025).
Why every AI answer needs a check
Every major assistant gets facts wrong often enough to matter, and a wrong answer reads exactly like a right one. ChatGPT Search signalled uncertainty only 15 times in 200 questions while getting 134 wrong in the Tow Center test (March 2025). In NewsGuard's August 2025 audit, 10 leading chatbots repeated false news claims 35% of the time. Tool-by-tool results are on is AI accurate?
The routine below works the same way for ChatGPT, Gemini, Claude, Copilot, Perplexity or any other tool. Run it on anything you will publish, submit, act on or repeat.
- Step 1
List the claims
Copy the answer and underline every statement it depends on: each fact, name, number, date, quote and cited source. Ignore the framing and the summary sentences; they are only as good as the claims under them.
Then sort the claims by stakes. A wrong date in background detail is a nuisance. A wrong dosage, legal deadline or citation in a submitted essay is a problem. Check the high-stakes claims first.
Free toolsa notes app or a printed copy and a pen. To see whether a claim has already been checked by professional fact checkers, search it in Google Fact Check Explorer, which finds published fact checks by text or image, or ask Snopes FactBot, which answers from the Snopes archive and itself tells users to check its cited sources.
- Step 2
Open every source
Click every link and confirm two things: the page exists, and it says what the AI says it says. This step catches the most damaging errors. Grok 3 produced 154 links to error pages in 200 answers in the Tow Center test, and Google AI Overviews gave answers not supported by their cited sources 56% of the time in the New York Times and Oumi analysis reported by Futurism (April 2026).
For a reference with no link, search the exact title. Papers turn up in Google Scholar or a library catalogue; a DOI resolves at doi.org; court cases appear in official court records or CourtListener. If a source cannot be found, treat it as invented. The full routine for references is on does ChatGPT make up sources?
Free toolsGoogle Scholar, your library catalogue, doi.org, CourtListener, and the Internet Archive's Wayback Machine for pages that have moved.
- Step 3
Check the numbers at the origin
Find every figure in its original source: the study, the official statistics release or the company filing. Not a news story about it, and not another AI summary. Check three things: the number itself, its unit (percent, percentage points, thousands, millions) and what exactly it counts.
Exact numbers are easy for a model to invent because they fit a familiar format. A figure that has no source attached is the first one to check.
Free toolsthe original report or dataset, official statistics sites for your country, and a search for the number in quotation marks with a key term beside it.
- Step 4
Check the dates
Ask whether the answer is still true today, and whether the dates in it are right. Models know least about recent events, and answers can describe a situation that has since changed. After the July 2024 assassination attempt on Donald Trump, Meta AI told some users it had not happened, and Meta later called those answers hallucinations (Meta, July 30, 2024).
Check the publication date of each source, and look for anything newer that supersedes it. For breaking news, go straight to news outlets rather than a chatbot.
Free toolsa news search with a date filter, the date line on each source, and the "last updated" note on official pages.
- Step 5
Match the quotes
Search every quote word for word, in quotation marks, and read it in its original context. AI assistants alter quotes, merge them and pin them on the wrong person. In the BBC's first study, 13% of quotes attributed to BBC articles were altered or missing from the original (BBC, February 2025). In the follow-up EBU/BBC study, Gemini had significant problems in 20% of its responses containing direct quotes.
If you cannot find the exact words in a primary source (a transcript, the original article, the published speech), paraphrase the idea with a real citation or drop it.
Free toolsa web search with the quote in quotation marks, the original transcript or article, and the Wayback Machine for pages that have changed.
Quick checklist
| Step | What to check | Done when |
|---|---|---|
| 1. Claims | Every fact, name, number, date, quote and source | You have a list, highest stakes first |
| 2. Sources | Each link opens and says what the AI claims | Every source is found and matches |
| 3. Numbers | Each figure, unit and definition at the origin | Each number traces to its original source |
| 4. Dates | The answer is current and its dates are right | Nothing newer contradicts it |
| 5. Quotes | Exact words in the original context | Each quote is found word for word |
When to go further
Some answers need more than five minutes. Go further for medical, legal and financial questions, for anything you will publish under your name, and for any claim that would change a decision. For those, confirm each key claim with a second, independent source, not a second chatbot, and ask an expert where one is available.
Dedicated tools can speed up long documents. Compare options on the best AI fact checkers. For a sense of which tools need the most checking, see the most accurate AI assistants and AI hallucination rates by model. Real cases where a skipped check went badly are on real AI hallucination examples.
A confident tone is never evidence: the check is what makes an AI answer usable.
Questions
How do I fact-check an AI answer quickly?
Check five things: the claims the answer depends on, the sources it cites, its numbers, its dates and its quotes. Open each source, find each number and quote in the original, and confirm the answer is still current. For most answers, this takes about five minutes with free tools such as Google Scholar and a web search.
Can I use another AI to fact-check an AI answer?
Use it to find leads, not to confirm facts. A second chatbot can repeat the same error or invent a different one. In NewsGuard's August 2025 audit, 10 leading chatbots repeated false news claims 35% of the time. Confirm each claim in a primary source you have opened yourself.
What part of an AI answer is most often wrong?
Sources are the most common problem in news answers. Sourcing issues affected 31% of answers in the EBU/BBC study, against 20% for accuracy issues. Links, quotes, exact numbers and recent events are the parts to check first, because they are easiest to get wrong.
Is there a free tool to fact-check AI?
Yes. Google Fact Check Explorer finds published fact checks of a claim, and Snopes FactBot answers from the Snopes archive. Neither judges a brand-new claim on its own, so pair them with Google Scholar for references and a quoted web search for quotes and numbers.
Do I need to fact-check Perplexity if it shows sources?
Yes. Perplexity had the lowest error rate of 8 tools in the Tow Center test, but still named the wrong source 37% of the time. A visible citation shows where to check, not that the claim is right. Open the source and find the claim inside it. More on is Perplexity accurate?
Sources
- Independent study
- Independent study
- Independent study
- News audit
- News audit
- Vendor claim
- Vendor claim
