In 2023 a New York law firm filed court papers citing six cases that did not exist. ChatGPT had invented them, quotes included. When the lawyer asked ChatGPT whether one of them was real, it said it did exist and could be found on Westlaw and LexisNexis, according to his own sworn statement. A federal judge ordered the firm and its two lawyers to pay a $5,000 penalty.
That is what a made-up AI answer looks like: calm, specific and wrong. OpenAI’s own research calls these “hallucinations”, meaning plausible but false statements, and said in September 2025 that they still happen in its latest models. Its explanation is that models are rewarded for guessing rather than for saying “I don’t know”. You can’t spot them by tone. You can catch most of them in five minutes.
The 5-minute checklist
Use it on any answer you plan to rely on, repeat, or send to someone else.
1. Mark what can be checked (30 seconds). Underline every name, number, date, quote and link. These are the parts a model most often invents. Anything left over is opinion or summary, and needs less checking. Anything you can’t check, treat as unconfirmed.
2. Open every source (90 seconds). Click each link. Does the page exist? Does it say what the answer says? For a book, report or case, search its exact title in a search engine or an official database. No source given? Ask for one, then check that source too.
Constructed example, not a real answer: “A 2024 study (Chen and Mills, Journal of Applied Cognition, vol. 12) found that 68% of people…” Search the exact title and authors. If no such paper turns up, drop the claim. A real-looking reference proves nothing until you have found it yourself.
Do not ask the same AI “is this real?” In the 2023 case, the lawyer did exactly that, and the chatbot assured him the case existed, as the court’s findings record. The judge also found that the case number and journal citation on one fake opinion belonged to other, real cases. A quick lookup shows that kind of mismatch.
A second real case, this time a wrong answer rather than an invented citation. In Moffatt v. Air Canada, an airline chatbot told a customer travelling after a family death that he could claim a bereavement fare after the trip. The chatbot’s own link led to a page saying the opposite. A British Columbia tribunal ordered Air Canada to pay $812.02 and called its suggestion that the chatbot was a separate entity responsible for its own actions “a remarkable submission”. Lesson: when the answer links to the official page, read the page, not the summary.
3. Redo the numbers (60 seconds). Check units, totals and percentages with a calculator. Find the original table or report.
Constructed example, not a real answer: “Revenue grew 40% to $12m, up from $9m.” Nine to twelve is 33%, not 40%. Either the figure or the percentage is wrong, and the whole answer is now suspect.
4. Check the dates and the word “latest” (45 seconds). A model can answer from old training data and still sound current. Words like “currently”, “latest” and “as of today” are warnings. Prices, laws, software versions, office-holders and opening hours change, so check the date on your source.
Constructed example: “The latest version is 3.2, released last month.” Open the maker’s release page. If it shows 3.4, the answer is stale.
5. Cross-check with one independent source (75 seconds). Find the key claim in a second place that isn’t the same AI rewording itself: the official site, the original document, a reputable news report. If two independent sources agree, relax a little. If you can’t find it anywhere else, don’t use it.
Timings: 30 + 90 + 60 + 45 + 75 seconds = 5 minutes.
Quick red flags
- A source with a perfect-sounding title that you can’t find anywhere.
- A quote with no page, date or speaker.
- Very round or very neat numbers.
- An answer that agrees with whatever you hoped to hear.
- More detail than your question could possibly justify.
- It insists it is right when you question it.
In the court case, when asked to “show me more cases”, the chatbot supplied more, according to the judge’s findings. Pressing for sources can produce more inventions, so check each one.
What to do when you find a mistake
- Stop using it. Don’t copy it into a report, email or post.
- Tell the AI exactly what’s wrong, for example “that link doesn’t exist”, and ask it to say what it can’t confirm. Treat the reply as a lead, not as proof.
- Go to the original source yourself and take the fact from there.
- If you already shared it, correct it quickly and openly. The judge in the 2023 case noted that if the lawyers had come clean early, “the record now would look quite different”.
- Check the other parts of the same answer. One invention often means more.
Even good answers need checking when the stakes are high
An answer that passes all five steps can still be incomplete or out of date. For anything about health, money, law or safety, or anything you’ll sign, publish or send, treat the AI as a first draft and confirm with the official source or a qualified professional. Our piece on people asking ChatGPT about their health every week shows how common the habit is. This guide is general information, not medical or legal advice.
Related on TSN: The Complete Beginner's Guide to AI: From Neural Networks to Real-World Applications; Natural Language Processing: The Complete Guide to Teaching Machines Human Language; 600,000 People Ask ChatGPT About Their Health Every Week – And 70% Do It After Hours; The Complete AI Glossary; How to Create Your Own AI Agents in Grok (Step-by-Step)
Sources
- OpenAI, “Why language models hallucinate”, 5 September 2025 (definition: plausible but false statements; still occur in newer models). https://openai.com/index/why-language-models-hallucinate/
- Judge P. Kevin Castel, Opinion and Order on Sanctions, Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y., 22 June 2023). https://caselaw.findlaw.com/court/us-dis-crt-sd-new-yor/2335142.html (docket: https://www.courtlistener.com/docket/63107798/mata-v-avianca-inc/)
- Civil Resolution Tribunal of British Columbia, Moffatt v. Air Canada, 2024 BCCRT 149, 14 February 2024. https://decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do
