Checking every fact in an AI draft before it goes live

AI tools produce wrong facts that look exactly like right ones. This chapter explains why, what it has cost others, and the method that catches the errors before a customer does.

Check every number, date, name, rule, price, quotation and source in an AI draft by opening the original source yourself and comparing the exact words. Google says generative models "don't retrieve facts, but predict a likely sequence of words", so their errors look as confident as their facts. Keep a list of each fact with its source and date, remove anything you cannot confirm, and recheck the list whenever the facts change.

Why AI tools invent facts

It helps to understand how the error happens, because then the method to catch it makes sense. Google's help page on generative AI explains it in one sentence: "generative models don't retrieve facts, but predict a likely sequence of words based on their training data." The page continues: "Because of this, generative AI outputs may contain inaccuracies (also known as hallucinations)."

Think about what that means in practice. When you ask a chatbot for the fee to apply for a building permit in your city, it predicts the kind of number that tends to appear in sentences like that. The number may be right, or it may be last year's number, or a number from another city. The sentence sounds equally sure in each case.

For this reason Google's conclusion is direct: "It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing." Notice the word "manually". The check is done by a person who opens the source.

What wrong facts have cost others

Two cases show the cost clearly, and both are public records that anyone can read.

In Mata v. Avianca, a case in the federal court in New York, two lawyers filed a brief that cited court decisions which did not exist. The judge's opinion of 22 June 2023 says they "submitted non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT, then continued to stand by the fake opinions after judicial orders called their existence into question." The court imposed a penalty of $5,000 on the lawyers and their firm. The same opinion is careful to add that "there is nothing inherently improper about using a reliable artificial intelligence tool for assistance." The fault was the missing check.

In Moffatt v. Air Canada, decided by British Columbia's Civil Resolution Tribunal on 14 February 2024, the chatbot on Air Canada's website told a grieving customer they could apply for a bereavement fare after travelling. The airline's own policy said otherwise. The tribunal ordered Air Canada to pay $812.02 and wrote: "It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot."

A small business is unlikely to face a federal judge over a blog post. But the principle in both cases applies to every website: the publisher answers for what it publishes, whoever or whatever wrote it.

The facts that go wrong most often

Some kinds of facts are riskier than others, because they change over time or differ from place to place. These are the ones to check first.

• Prices and fees, especially government fees, which change every year or two.
• Laws, regulations and building codes, which differ from one city, province or state to another and are amended often.
• Statistics, where a model may combine a real figure with the wrong year, place or source.
• Quotations, which a model may paraphrase and then put inside quotation marks as if they were exact.
• References and links, including titles of studies, court decisions and articles that sound real and do not exist.
• Facts about named companies and people: awards, memberships, licences, years in business, which a model may repeat from the company's own marketing as if they were independent facts.
• Dates and deadlines, which can be right for one year and wrong for the next.

The method, step by step

The method is simple, and the discipline lies in doing every step for every fact.

• Underline every claim in the draft that a reader could check: numbers, dates, names, rules, prices, quotations, statements about other businesses.
• For each one, find the primary source: the body that publishes the figure or makes the rule. A city fee comes from the city's fee schedule, a law from the official text, a statistic from the agency that collected it.
• Open the source yourself and find the exact figure or sentence. A search result summary is a pointer to a source, and a different thing from the source.
• Compare word for word. Check the number, the unit, the year, the place and any conditions attached ("for buildings over three storeys", "from 1 January").
• Record the fact in a list kept with the page: the claim, the source, the link and the date printed on the source.
• If you cannot confirm a fact, remove it or rewrite the sentence without it. A page with one fewer number is better than a page with a wrong one.
• Put the date on the page where it matters: "fees as published by the city on 1 March 2026".

Where to check each kind of fact

Facts about other businesses need extra care

One kind of error deserves its own warning, because it can harm a third party. A chatbot asked about a competitor, a supplier or a partner may repeat what that company says about itself: its awards, its memberships, its years in business, its number of projects. A company's own website is a statement by the company. It proves what the company claims, and it proves nothing more.

When a page names another business, confirm each claim with the body that would know: the association's public member list for a membership, the awarding body's list of winners for an award, the licensing registry for a licence. If no independent source confirms it, either say whose claim it is ("the company says it has completed 200 projects") or leave it out.

Health, money and safety pages

Google holds some topics to a higher standard. Its help page on people-first content says that for topics "that could significantly impact people's lives or well-being (YMYL topics), the content must be highly accurate and consistent with established expert consensus." YMYL stands for "your money or your life": health, finance, legal matters, safety.

On these pages, have a qualified person check the facts as well as the editor: a registered health professional, an accountant, a lawyer, an engineer. Name that person on the page, with their qualification. Google's 2023 guidance notes that on topics like "health, civic, or financial information", its "systems place an even greater emphasis on signals of reliability."

What a fact list looks like

The fact list does not need special software. A simple table in a document or a spreadsheet, saved next to the page, is enough. Each row holds one claim from the page and four columns: the claim as it appears on the page, the source's name, the link, and the date printed on the source. Add a fifth column for the date you checked it.

This guide was built the same way. Every figure and quotation in its chapters was checked against the sources listed at the end of each page, which were opened on 4 November 2026, and the list of those sources, with the exact sentences used, is kept with the guide so that it can be rechecked when Google or a court publishes something new.

Keep the fact list alive

Facts age. A fee changes on 1 January, a rule is amended, a statistic is updated. The list of facts you built for a page is what makes updating it possible, because it tells you exactly what to recheck and where.

Set a reminder to recheck each page's fact list at least once a year, and sooner for pages with government fees or rules. When something changes, update the page, update the list, and only then change the date on the page. Google lists changing dates "to make them seem fresh when the content has not substantially changed" among its warning signs.

Part of a larger guide

This page is one part of Avoiding AI slop. The other parts:

• What AI slop is
• What Google says
• Signs of AI writing
• The editing process
• First-hand experience
• Bylines and disclosure
• Publishing at scale

Questions people ask

Why does AI make up facts?

Because language models predict likely words from their training text. Google's generative AI help page says the models "don't retrieve facts, but predict a likely sequence of words based on their training data", so their output "may contain inaccuracies (also known as hallucinations)." A wrong figure comes out in the same confident sentence as a right one, which is why every figure needs checking.

What is an AI hallucination?

It is a statement produced by an AI tool that is false but presented as fact, such as a fee that does not exist, a quotation nobody said or a study that was never published. Google's help page uses the word for the inaccuracies generative models can produce. Hallucinations look exactly like correct text, so only checking the source reveals them.

What happened in the Mata v. Avianca case?

Two New York lawyers filed a brief citing court decisions invented by ChatGPT, and kept defending them after the court questioned whether they existed. In an opinion dated 22 June 2023, Judge Castel imposed a $5,000 penalty on them and their firm. He also wrote that using a reliable AI tool for help is acceptable, and that the duty to check accuracy stays with the user.

Is a business liable for what its chatbot or AI content says?

In Moffatt v. Air Canada, British Columbia's Civil Resolution Tribunal ruled on 14 February 2024 that it was. Air Canada's chatbot described a fare rule wrongly, and the tribunal ordered the airline to pay $812.02, writing that it "makes no difference whether the information comes from a static page or a chatbot." Rules differ by country, so ask a lawyer about your own situation.

What counts as a primary source?

A primary source is the body that creates or publishes the fact: the city for its fees, the legislature or regulator for a law, the statistics agency for its data, the court for its decision, the author for a quotation. A news article or a blog that reports the fact is a secondary source. Check the primary source whenever it is available.

How do I check facts about competitors or partners mentioned on my site?

Confirm each claim with an independent body that would know: an association's public member list, an awarding body's published winners, a licensing registry. A company's own website shows only what it says about itself. If nothing independent confirms a claim, attribute it ("the company says") or remove it, because a false statement about a named business can harm it.

Can I trust the sources an AI tool gives me?

Not without opening them. AI tools can give titles, authors and links for sources that do not exist, or real sources that do not say what the tool claims. Open every link, find the exact sentence or figure, and confirm it matches. A source you have not opened is a guess about a source, and it should not appear on your page.

How should I show sources on my page?

Name the source in the sentence where a reader needs it ("according to the city's 2026 fee schedule"), and list the sources with their dates at the end of longer pages. Google's help page on people-first content asks whether content shows "clear sourcing". Dates matter as much as names, because they tell the reader how current each figure is.

How often should I recheck facts on published pages?

At least once a year for every page, and more often for pages with government fees, laws or prices, which change on fixed dates. Keep the fact list with the page so you know exactly what to recheck and where. Change the page's date only after you have updated the content, since Google treats date changes without real updates as a warning sign.

What should I do if I cannot confirm a fact in an AI draft?

Remove it, or rewrite the sentence so it no longer needs that fact. A page with one fewer number is better than a page with a wrong one. If the fact matters to the reader and no source confirms it, say what is known and what is not, or point the reader to the body that can give the current answer.

Sources

1. Google Search's guidance on using generative AI content on your website, Google Search Central. Last updated 2026-10-01, accessed 4 November 2026.
2. Mata v. Avianca, Inc., Opinion and Order on Sanctions, 22-cv-1461 (PKC), U.S. District Court for the Southern District of New York. Filed 22 June 2023, accessed 4 November 2026.
3. Moffatt v. Air Canada, 2024 BCCRT 149, Civil Resolution Tribunal of British Columbia. Issued 14 February 2024, accessed 4 November 2026.
4. Creating helpful, reliable, people-first content, Google Search Central. Last updated 2026-10-05, accessed 4 November 2026.
5. Google Search's guidance about AI-generated content, Google Search Central Blog. Published 8 February 2023, accessed 4 November 2026.

Keep reading

• How to avoid AI slop: the full guide
• The editing process from AI draft to published page
• What Google says about content written with AI
• Bylines, disclosure and AI
• Run the free Get Found Check

Want pages that sound like your business and bring enquiries?

Licheo's agents write and update your pages from your own facts, and a person checks every page before it goes live. Start with the free Get Found Check to see how customers find you today, or contact us for pricing.