AI answers have started coming back with citations: small numbers next to sentences, links to sites that exist. The links do a job, and it's a smaller job than it looks. A citation proves a source exists, not that the answer matches it. The check that finishes the work takes about a minute per claim once you know the four moves.
The four moves
Name the claims you plan to use. An AI response can run ten paragraphs, and checking every sentence is a slow way to read. The claims that need fact-checking are the ones you'd repeat, publish, or act on: the statistic in your report, the price in your spreadsheet, the event date in your newsletter. The rest is reading. For help deciding which claims are high-stakes, How to Verify AI Output: A Simple Framework sorts decisions by the cost of being wrong.
Open the source before you trust the citation. Click the link and find the sentence the claim rests on. Three failure modes to look for: the source doesn't exist, the source says something different, or the source is too old to support a claim about today. The mismatch is common enough that researchers have measured it. A 2025 study led by the BBC and the European Broadcasting Union had professional journalists evaluate more than 3,000 AI responses about news from ChatGPT, Copilot, Gemini, and Perplexity. 31% had sourcing problems: missing, misleading, or incorrect attributions. OpenAI's own help page gives the same advice: open a cited source and check that it supports the answer.
Search the claim itself. Take the key sentence, not the whole answer, and search it in a search engine inside quotation marks. Exact-phrase search can point you toward whoever said it first. For a statistic, keep going until you reach the original study or report, not a blog that cited another blog. Two details decide whether the number survives: the date, because a true figure from 2023 presented as current is still a wrong claim, and the population studied, because a survey of one country's drivers is not a worldwide number.
Drop what you can't confirm. If the search turns up nothing, the claim doesn't go in the report, the email, or the brief. A useful-sounding claim is still unconfirmed.
A check that passes: the 31% figure
We checked the study behind the second move while writing this post. Search a line like "31% of responses showed serious sourcing problems" and the results point to the BBC's announcement of the study from October 2025. Read it and the details confirm the claim: journalists at 22 public broadcasters in 18 countries evaluated more than 3,000 responses from ChatGPT, Copilot, Gemini, and Perplexity, across 14 languages. The number survived: the sample existed, the date was recent, and the wording matched.
Checking has a quiet payoff: you end up holding the original source instead of a claim about a source.
The claim-type cheat sheet
- A statistic: find the original study or report, then check the date and who was studied. If the answer names no source, treat the number as a placeholder until you find one.
- A quote: search the exact wording in quotation marks. If a person supposedly said it, look for it in a transcript, the organization's own website, or news coverage that recorded it.
- A citation (paper, case, article): search the title in Google Scholar or the publisher's database. If it doesn't appear, it probably doesn't exist.
- An event or date: find the primary announcement, the company press release, the government page, the court filing. Secondhand retellings can drift.
- An image: run a reverse image search to find where it came from. If it claims to show a documented event, look for the same photo from news organizations.
What skipping the check cost one law firm
In 2023, two New York lawyers filed a brief in the case Mata v. Avianca citing court cases gathered with ChatGPT's help. Six of the case citations didn't exist. When opposing counsel couldn't find the cases, one of the lawyers asked ChatGPT whether they existed, and it said yes, naming legal databases where they could supposedly be found. In June 2023, a federal judge sanctioned the lawyers and their firm $5,000.
The fabricated citations had case names, volume numbers, page numbers, and quotes. The lawyers hadn't run the check before filing. A search of the case names in a legal database could have exposed the problem.
FAQ
Are the citations in AI answers reliable?
The links often exist, and a link that exists isn't the same as support. In the BBC-led study of more than 3,000 responses, 31% had sourcing problems, including misleading or incorrect attributions. A citation tells you where to check. It doesn't do the checking.
How do I check if a quote from AI is accurate?
Search the exact wording in quotation marks. If a person supposedly said it, look for it in a transcript, the organization's own website, or news coverage that recorded it. If you can't find the phrase in a reliable source, treat it as unconfirmed.
Does ChatGPT's search mode fix hallucinations?
Search mode gives the model sources to work from, which reduces answers invented from memory. It doesn't check them for you: the BBC-led study tested ChatGPT, Copilot, Gemini, and Perplexity, all of which search the web, and still found 45% of responses with at least one significant issue. Search gives you something to click. You still have to check it.
Can I ask the AI to check its own answer?
You can, and it can catch slips, like a number that doesn't match the source right above it. It isn't verification: a model can be just as confident about its own mistakes, and it shares the same blind spots from one answer to the next. Outside sources are what settle it.
Want to go further?
- How to Verify AI Output: A Simple Framework covers which claims are worth checking and how much effort each stake level deserves.
- How to Use AI for Research Without Getting Burned covers sourcing and the limits of AI-generated citations for research work.