To fact-check AI writing, copy every checkable claim out of the draft into a separate list, then confirm each one in a primary source that you open and read yourself: the judgment, the statute, the original study, the company's own documentation. Anything you cannot trace gets rewritten in general terms or deleted. Never ask the same AI tool to confirm its own facts, and treat references, quotations and numbers as the highest-risk parts of any draft, because those are where the most expensive public mistakes have happened.

Fact-checking AI writing at a glance

  • Separate the claims from the prose. Names, dates, figures, quotes, legal points, product features and causal statements all go on a list before you edit a single sentence.
  • Go to the original. A summary, a news story or a chatbot's citation is a pointer, not proof.
  • Check that sources exist and say what the draft says. Fabricated references and real references that do not support the claim are both common.
  • Match quotations word for word. If you cannot find the exact wording, paraphrase with attribution or cut it.
  • Read numbers in context. Who was measured, when, where and how.
  • Keep a record. A short log of each claim, its source and the date checked makes corrections faster and protects you later.

Why AI drafts need a different kind of check

A human writer who gets something wrong usually leaves clues: hedged wording, a vague attribution, a gap where the research ran out. Language models tend to do the opposite. In a September 2025 paper, Why Language Models Hallucinate, researchers from OpenAI and Georgia Tech argue that models "sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty", because the way they are trained and scored rewards a confident guess over an honest "I don't know". The paper opens with an example: asked three times for one of the authors' birthdays, a leading open model gave three different wrong dates.

The practical consequence is that fluency tells you nothing about accuracy. Broadcasters have measured how often this goes wrong with real questions. In research published in February 2025, the BBC put 100 news questions to ChatGPT, Copilot, Gemini and Perplexity and had its own journalists review the answers. The BBC's report on AI assistants found that 51% of answers had significant issues, 19% of answers citing BBC content introduced factual errors in statements, numbers or dates, and 13% of quotes attributed to BBC articles were altered or did not appear in the article cited.

A follow-up led by the European Broadcasting Union, involving 22 public service media organisations in 18 countries and 14 languages, reviewed 3,062 responses generated in late May and early June 2025. Its News Integrity in AI Assistants report found that 45% of responses had at least one significant issue. Sourcing was the biggest single problem (31% of responses), followed by accuracy (20%). The same failure types turn up in AI-drafted articles and reports.

What goes wrong in practice

The well-documented failures fall into a small number of patterns, and each one suggests a specific check.

Invented sources that the tool then vouches for. In Mata v. Avianca in New York, lawyers filed a brief citing cases that ChatGPT had made up. When they asked the chatbot whether the cases were real, it assured them they could be found in legal databases. Judge P. Kevin Castel fined the lawyers and their firm $5,000 on 22 June 2023, as the Mata v. Avianca case summary records. The lesson for editors is simple: a model's confirmation of its own output is not verification.

Real-looking references that do not exist or do not say what is claimed. In England, the Divisional Court's judgment of 6 June 2025 in Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank dealt with two cases of suspected AI-generated citations. In Al-Haroun, 45 citations were put before the court and 18 of the cases did not exist; many of the rest did not contain the quoted passages or did not support the point. Dame Victoria Sharp, President of the King's Bench Division, wrote that such tools "may cite sources that do not exist" and "may purport to quote passages from a genuine source that do not appear in that source".

Plausible policies that a business never had. In February 2024 a Canadian tribunal ordered Air Canada to pay a customer $812.02 after its website chatbot said a bereavement fare could be claimed retrospectively, which the airline's actual policy did not allow. The tribunal decision in Moffatt v Air Canada rejected the idea that the chatbot was responsible for its own actions: "It should be obvious to Air Canada that it is responsible for all the information on its website."

Fabricated content published without a read-through. In May 2025 a syndicated summer supplement in the Chicago Sun-Times recommended books that do not exist, attributed to real authors. The Sun-Times account of what happened quotes the freelance writer admitting he had republished a list an AI program produced; NPR's report traced how the same material reached other papers.

Errors in a professional report. In October 2025 Deloitte Australia agreed to refund the final instalment of an AU$440,000 contract after a University of Sydney academic found fabricated references and an invented court quotation in a report for the Department of Employment and Workplace Relations. The revised version disclosed the use of a generative AI model, according to The Register's report on the Deloitte refund.

How to fact-check an AI draft step by step

Extract the claims first

Read the draft once for sense, then go through it line by line and copy every statement a reader could challenge into a table: the claim, where it appears, the source you will check and the result. Include the quiet claims as well as the obvious ones, such as job titles, the year a law came in, "the first", "the largest" and any "because". Do this before editing for style, so you do not polish a paragraph you will have to delete.

Find the primary source for each claim

Work from the original document wherever one exists. For UK law, go to legislation.gov.uk for statutes and Find Case Law at The National Archives for judgments; the Ayinde judgment names both kinds of database among the authoritative sources lawyers should use. For UK statistics, start with the Office for National Statistics or the department that published the figure. For a research paper, confirm it exists and has the authors, year and title the draft gives, for example through Google Scholar or the publisher's site, then read the relevant section rather than the abstract alone. For product features, prices and limits, use the vendor's own current documentation and note the date you checked it.

Check that the source supports the sentence

Existence is only half the test. Open the source and find the passage. AI tools often stitch together two true things into one false claim: a real company with another firm's launch date, a real report with a figure from a different year, a real judge with a quotation from another case. Mike Caulfield's SIFT method (Stop, Investigate the source, Find better coverage, Trace claims, quotes and media to the original context) was written for social media, but its last step is exactly what this stage needs.

Match quotations exactly

Search for a distinctive phrase from the quotation in the original source. If the words differ, use the original or paraphrase with clear attribution. If you cannot find the quote at all, delete it. The BBC's finding that 13% of quotes were altered or absent is a good reason to treat every AI-supplied quotation as unverified until proven otherwise.

Test every number

For each figure, write down what was measured, the population, the place, the date and the method. A percentage with none of these is not usable. Watch for round numbers with no source, figures that have since been updated, and UK articles quietly quoting US or global data. If the number cannot be traced, replace it with an accurate general statement or drop the point.

Warning signs that a draft needs closer scrutiny

The judiciary of England and Wales publishes a list of signs that material may have been produced by AI. The AI Guidance for Judicial Office Holders, dated 31 October 2025, is written for courts but transfers well to any editorial desk. It says AI tools may "make up fictitious cases, citations or quotes, or refer to legislation, articles or legal texts that do not exist", and it flags warning signs including:

  • references that do not sound familiar or have unfamiliar citations;
  • American spelling or overseas examples in material meant for a UK audience;
  • content that "appears to be highly persuasive and well written, but on closer inspection contains obvious substantive errors";
  • a leftover prompt or a stray "as an AI language model" phrase.

To those, editors can add a few of their own. Be suspicious of phrases such as "studies show" or "experts agree" with no named study or expert, comparisons with no baseline ("faster", "cheaper", "safer" than what), and confident statements about what a tool integrates with, supports or complies with. Product and compliance claims should only stay in if the vendor's documentation confirms them; otherwise soften them to "check the current documentation" or remove them.

Tools that help, and their limits

Tools speed up the work but do not replace reading the source.

  • Exact-phrase search. Searching a quotation in double quotation marks is still the quickest way to find its origin.
  • Fact-check databases. Google Fact Check Explorer searches fact checks published by fact-checking organisations, which is useful for viral claims and statistics that have already been examined. Google's documentation says it is phasing out ClaimReview markup in Search while the Explorer tool continues to support it (checked October 2026).
  • Source buttons in assistants. Some chat tools show where an answer came from. Gemini's help page on viewing sources explains that a Sources button appears "when sources are available", and that not every response has one (checked October 2026). A linked source is a starting point: the EBU study found sourcing was the most common failure.
  • UK fact-checkers. Full Fact publishes checks on claims in UK public debate, which can save time on contested political or health statistics.

Asking a second AI model to check the first can surface questions worth investigating, but it carries the same risk the Mata lawyers ran into. Treat its answers as leads to follow up in primary sources.

Building the check into your workflow

Fact-checking works best when it is a fixed stage with an owner, not something squeezed in before publication. A simple routine:

  1. Draft with AI if it helps, but ask it to mark every factual claim and give a source for each, so the checker's list is partly built.
  2. Extract claims into a log with columns for claim, source URL, result (confirmed, corrected, removed) and date checked.
  3. Verify against primary sources, starting with names, dates, figures, quotes, legal points and product details.
  4. Correct or cut anything unsupported, and add links where they help readers check for themselves.
  5. Have a second person spot-check the riskiest claims on regulated, legal, health or financial topics.
  6. Re-check time-sensitive facts such as prices, fees and rules whenever the piece is updated, and show the date checked in the text.

The time this takes depends on how many claims the piece makes rather than its length. A 1,500-word opinion piece may have a dozen checkable statements; a guide to a regulation may have fifty. Build the check into deadlines from the start, and keep the log with the published article so later updates can start from what was already confirmed. More how-to guides on writing with AI are on AI Disraeli.

Frequently asked questions

Can I ask the AI tool to fact-check its own writing?

You can ask it to list its claims and suggest sources, which saves time. You cannot rely on its verdict. In Mata v. Avianca the chatbot confirmed that invented cases were real. Every claim still needs a human to open the source.

Which claims should I verify first?

Start with whatever would cause the most harm if wrong: legal and regulatory statements, health and financial information, quotations attributed to real people, statistics and named references. Then move to names, titles, dates and product details.

Is a citation in an AI answer enough?

No. Check that the cited page exists, that it is the original rather than a summary, and that it actually contains the statement. The Ayinde judgment and the EBU study both show real-looking citations that did not support the claim.

What should I do with a claim I cannot verify?

Rewrite it in general terms that you can support, or remove it. Do not publish it with a vague attribution such as "reports suggest" in the hope that it is true.