To check whether ChatGPT made things up, run its answer through seven checks before you use it: mark each specific claim, then open its source and check numbers and quotes yourself. This page gives the checklist and a 20 minute audit of last week's work.
The checklist comes from the book Don't Replace Me by Dmitry Kargaev, who also wrote this page.
What happened in Mata v. Avianca
In the book's account of Mata v. Avianca, lawyer Steven Schwartz cited cases that did not exist. In May 2023, Judge P. Kevin Castel called six of them "bogus judicial decisions with bogus quotes and bogus internal citations." On 22 June 2023 he fined Schwartz, the signing lawyer and the firm $5,000 and found bad faith.
The book's lesson is that your name is on the work, so you verify it. It also cites the Charlotin database, which stood at 2,266 cases when the author checked in early November 2026.
Why ChatGPT sounds sure when it is wrong
OpenAI's own help page says ChatGPT can give incorrect or misleading output while sounding confident. Its list of hallucinations includes fabricated quotes, studies, citations and references to non-existent sources.
The book gives one reason, taken from a September 2025 OpenAI paper: the tests used to score models reward a confident guess over "I don't know". It adds that AI is strongest where something can check the answer, and elsewhere it guesses confidently. In a work document, the thing that can check the answer is you.
Asking the same model "are you sure?" does not change that. In the book's words, it "feels like checking, but it's the same guesswork running a second time." The book also says the errors have become rarer but subtler, such as a real link attached to the wrong claim, and that they look finished faster.
Before you start
You need the ChatGPT answer pasted into a document you can mark up, access to the original sources it mentions, a spreadsheet, and any notes or files of your own on the topic.
To see whether ChatGPT looked anything up, check whether the search or deep research icons are enabled in the input bar, or whether the response cites web sources. Depending on your plan, ChatGPT may have Search and Deep research, which give cited answers. Without them, the answer comes from training data alone, and events after the training cutoff are missing.
This came from a book.
Don't Replace Me
200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff.
Get the Book →The seven-step check
- Underline every specific claim. Copy the answer into a document and underline each name, date, number, quote, citation and "according to" statement. You should end with a marked copy in which every checkable sentence is underlined.
- Open every source and find the supporting sentence. Click each link or look up each citation yourself, and paste the sentence that backs the claim next to it. Expect one pasted sentence per claim; a claim without one is unsupported.
- Test each key claim a second way. Confirm it by a route other than the one that produced it, such as the original report, your own records or a colleague who knows. Expect a confirmation that does not come from the same chat.
- Redo the math in a spreadsheet. Enter the raw figures yourself and recompute every total, difference and percentage in the answer. Expect either matching results or a list of numbers to correct.
- Match every quote word for word. Find the quoted line in the original text and compare the two. Expect an exact match; if you cannot find the line, cut the quote.
- Read it as the recipient. Reread the answer from where the person receiving it sits, and note every point they would question or already know. Expect a short list of questions to answer before you send.
- Name the most embarrassing possible mistake and check that one again. Decide which single error would hurt most, such as a wrong client name or a wrong figure, and check it against the original record. Expect to have verified the item you would least want wrong.
Two habits for the prompt itself
Give it your own material. The book says AI works with what you give it and leans toward what you seem to want, so it suggests briefing ChatGPT like a smart intern. Paste in your notes, figures and documents, say who the audience is and what you want, list what to avoid, and tell it to ask up to three questions and to say when it is unsure. Chapter 12 puts the difference this way: "Same machine. The second prompt has about ten things in it that only you knew."
Then ask for the three weakest points. Once the draft exists, send a second message that has ChatGPT play the toughest reader and list the three weakest points of its answer and every fact it is unsure of. Use the list to decide which claims to open first in step 2. The steps above still do the checking.
Audit last week in 20 minutes
The book ends its chapter on verification with a 20 minute exercise called "audit last week". Here is a version for this page. Pick one thing you sent last week that ChatGPT helped write, and run steps 1 to 7 on it as if you were about to send it today.
You will come away with either a clean pass or a list of errors that are already out. If you find one, the book's chapter 8 gives a script for the correction: say what was wrong, give the correct version, say that it is yours, say how you checked the rest, and say what changes. It also advises mentioning the tool once as a fact, never as the culprit.
If it does not work
A link opens but nothing on the page supports the claim. Treat the claim as unsupported and cut it, or rewrite it from what the page does say.
You cannot find the source at all. OpenAI lists references to non-existent sources among ChatGPT's hallucinations, so remove the citation and the claim resting on it unless you can support it another way.
The claim is recent. Models are trained up to a cutoff, and unless search or deep research is used, their responses do not include later events. Look the fact up in a current source.
ChatGPT could not read the page you wanted it to use. OpenAI says it may fail to get information from a website because of technical issues, paywalls or robots.txt preferences. Open the page yourself and paste in the relevant text.
Real cases of invented facts at work
| Case | Date | Outcome |
|---|---|---|
| Mata v. Avianca | 22 June 2023 | $5,000 fine (Schwartz, the signing lawyer, the firm) |
| Jake Moffatt and Air Canada's chatbot | February 2024 | Tribunal member Christopher Rivers rejected Air Canada's argument that the bot was "a separate legal entity that is responsible for its own actions"; about CA$812 awarded |
| Morgan & Morgan, hoverboard suit against Walmart | February 2025 | A Wyoming federal judge fined the firm's lawyers |
Chapter 8 of the book says: "A model can't be fired, sued or embarrassed, and the client knows it."
Where the checklist comes from
The seven steps are the verification checklist in chapter 14 of Don't Replace Me: A Survival Guide to the AI Apocalypse, titled "It Will Lie to Your Face". Its rule is "verify before you sign". The two prompt habits come from chapter 12.
The second edition is on Amazon as a Kindle ebook at $2.99, a paperback at $14.99 (174 pages) and a hardcover at $24.99. The introduction is a free download from the book's site. There is nothing to buy after the book: no course and no community.
