Give it the facts before it writes, and check the claims after. An AI tool writing about your company without your approved facts in front of it will fill the gaps with what’s probable, and what’s probable is often wrong: a plausible price, a client you never had, a figure rounded to something impressive. The only reliable defence is a short record of what’s true, with sources, that the tool reads every time, and a check of each draft against it before anyone sends it.
How often it happens
The BBC asked ChatGPT, Copilot, Gemini and Perplexity 100 questions about the news in December 2024 and had its journalists mark the answers. Over half (51%) had significant problems. Of the answers that cited BBC content, 19% introduced factual errors such as wrong dates and numbers, and 13% of the quotes had been altered or didn’t exist in the article cited. (Press Gazette, The Register)
The follow-up, run by the European Broadcasting Union with 22 public broadcasters in 18 countries, checked more than 3,000 answers in 14 languages. 45% had at least one significant issue, 31% had a serious sourcing problem, and 20% had a major accuracy problem, meaning invented or outdated detail. (EBU, Reuters)
Columbia’s Tow Center gave eight AI search tools 1,600 excerpts and asked each to name the article and link to it. More than 60% of the answers were wrong, and the tools were confidently wrong far more often than they declined to answer. (Columbia Journalism Review, Nieman Lab)
Those studies test tools answering about the news, where the facts are public. The tools have far less to go on about your company.
What it costs when it’s your company
In February 2024 a Canadian tribunal ordered Air Canada to honour a bereavement refund that its website chatbot had invented. The airline argued the chatbot was a separate legal entity responsible for its own actions. “It should be obvious to Air Canada that it is responsible for all the information on its website,” the tribunal said, and awarded C$812. (The decision, CBC)
In April 2025 the AI support agent at Cursor, a coding tool, told a customer that logging out when switching machines was deliberate, citing a one-device-per-subscription policy. There was no such policy. The co-founder apologised on Reddit, and several customers said they were cancelling. (The Register, CX Today)
In October 2025 Deloitte Australia repaid part of a A$440,000 fee after a report for the Australian government was found to cite academic papers that didn’t exist and quote a Federal Court judge who’d never said it. The corrected version disclosed that an Azure OpenAI model had been used. (AP via BNN Bloomberg, TechSpot)
Each of those was a model filling a gap with something plausible, and a person trusting the result.
What to do
Write down the facts you’d defend. What you do, who for, the figures you can back and where each came from, the clients you may name, your prices, your credentials. Date every line. Keep it short enough that someone will maintain it.
Put that record where the AI tool reads it every time, and tell the tool to use only those facts and to say when something isn’t in them. How to do that in Claude and ChatGPT is covered in their own guides.
Then check. Read each draft against the record before it goes out, claim by claim. The figure, the client name, the date, the superlative. A model’s confidence tells you nothing about its accuracy, and the Tow Center study found the paid tiers were often more confidently wrong than the free ones.
What people get wrong
They ask the model to fact-check itself, and it says yes. They upload the whole website as the record, which includes every stale figure that was ever published. And they treat a wrong fact as a one-off rather than a sign that the record and the check are both missing.
How BeigeCheck does it
BeigeCheck holds your company’s approved facts in six areas, each with its source and version, and gives Claude, ChatGPT and Microsoft 365 Copilot the ones a draft needs before they write. Nothing is added or changed without a proposal and an approver.
After the draft is written, the Beige-ometer checks every claim in it against those facts and lists the ones they don’t support, so a figure that’s wrong or out of date is caught in the chat, before it’s sent. It checks each draft against your own approved sources rather than the open web. The facts are yours, you approved them, and the tool has nothing else to draw on.