AI Invents Figures, Sources and Terms: How to Spot It and Protect Yourself
By Anis Hammouche·October 22, 2026
You ask an AI to draft the offer page for your freelance work or small agency, and it slips in a percentage, the name of a study and a legal article. It all looks plausible, none of it is checked, and one of the three may be invented. That is a hallucination: a false answer presented as true. Its tone says nothing about whether it is right, so you go back to the source before you publish or send.
A real case
In 2023, in the United States, two lawyers filed a brief before a federal court in New York, in the case Mata v. Avianca, that contained invented court decisions. They had obtained them from ChatGPT (CNBC, 30 May 2023). The citations, the quotations and the internal references were false. Six of these decisions turned out to be invented, and the judge fined the two lawyers and their firm $5,000 (Legal Dive, 26 June 2023). The trap affects any professional who copies a source without opening it.
Why it invents
A language model learns to predict the next word. It is very good at producing text that looks like an answer. It does not check every statement against a base of facts at all times.
In a paper published in September 2025, Adam Tauman Kalai, Ofir Nachum, Santosh Vempala and Edwin Zhang explain that models hallucinate because their training and evaluation procedures reward guessing over admitting uncertainty. They compare it to candidates who guess on hard exam questions, because guessing improves the score (Kalai et al., arXiv, OpenAI). They propose changing how existing benchmarks score answers.
For you, the result is this: the AI has a natural pull towards a complete answer, even when it lacks elements.
Spotting a hallucination
No single signal proves that an answer is wrong. But several together should alert you.
| Signal | Why it is suspicious | What to do |
|---|---|---|
| A precise figure with no source | Precision without origin is often filled in | Find the body that publishes it |
| A source with no link | A plausible title proves nothing | Open the page or document |
| A law article cited from memory | Article numbers are easy to invent | Check on Légifrance |
| A study you cannot find | An author name or title can be fabricated | Search the exact title in a search engine or academic database |
| An answer that is too perfect | It matches your question exactly | Ask what has not been verified |
An example of checking
Imagine you are preparing the offer page for your business and want to write that a large share of French companies use AI. An AI answer might give you a percentage. Before you publish it, go to the source. INSEE reports that 18% of companies with ten or more employees in market sectors (excluding agriculture, finance and insurance) said they used at least one AI technology in 2025, according to Insee Première no. 2120, published on 21 July 2026 (INSEE).
This figure is usable, provided you cite it correctly. It covers "at least one AI technology", not generative AI alone. An AI that had given you a figure on ChatGPT usage would not have provided the same information. This is the kind of drift that checking the source prevents.
The terms follow the same rule
A delivery time, a warranty, a price or a commitment period can be invented in the same way. The reflex is identical: you only repeat a term if you find it in the contract, the vendor's official page or your client's document. For a quote, this point is covered in detail in the article on checking prices and terms.
Protecting yourself
- Go back to the original source. Open the document. Read the passage that contains the information. A cited source that does not say what you claim it says is a hallucination.
- Favour the data producers. INSEE for French statistics, Légifrance for legal texts, official sites for rates and obligations.
- Ask the AI to flag its doubts. For example: "List the sources, and say for each whether you verified it. If you are not sure, write 'I do not know'." This reduces the problem without removing it.
- Never give a reference you have not opened. A link you have not read proves nothing, even if the page exists.
- Treat sensitive areas separately. Legal, tax, health, safety: an AI answer is a starting point, not a conclusion.
What checking does not fix
Checking sources protects against invented facts. It does not protect against omissions. An answer can cite an exact source and forget the other half of the subject, or pick a study that leans towards your question.
That is why you also ask for what is not said: "what reservations or counter-arguments exist?". You do not get a complete truth, but you limit confirmation bias, the habit of finding only what you look for.
For an important decision, cross-check two sources of different kinds, as you do when handing work to an AI the way you would to an intern: an official body and feedback from the field, for example a client or a peer. A single source, even an accurate one, is still a single source.
Frequently asked questions
Do the latest models still hallucinate?
Yes. Recent models make fewer errors on some questions, but the risk does not disappear. Checking is necessary whatever the tool.
Is an AI with web search reliable?
It can find real sources, but it can also summarise a page wrongly. Always open the cited source and check that the passage matches.
How do you tell a client that information is wrong?
Show the original source and the exact passage. The most effective approach is to correct without dramatising: the goal is accurate information, not a debate about the tool.
Stuck on a specific case? Write to me via the contact page.
