You ask your assistant a health question. It answers with confidence, and adds: "According to a study by a major American university published in 2023, this phenomenon affects 30% of adults."
The sentence has everything it needs to inspire trust: a prestigious university, a date, a precise figure. One question remains: does this study actually exist?
Why an AI invents sources
An AI does not search a library. It produces the most plausible text to answer your question. And a reference to a study is a very common pattern in the serious texts it has read. It can therefore fabricate a perfectly credible reference that never existed. We explained this mechanism in our article on hallucinations.
Even when the assistant searches the internet, it can cite a real page while making it say more, or something other, than what it says.
Five checks, from the quickest to the surest
1. Click the link. If there is one, open it. A page that cannot be found is the first warning sign.
2. Search for the exact title. Copy the title of the study or article into a search engine. For scientific work, specialised engines exist, such as Google Scholar. If nothing matches, be suspicious.
3. Check that it really says that. This is the step most often skipped, and the most important. Very often, the source exists, but it talks about something else, with different figures or a more cautious conclusion. Look on the page for the passage that matches.
4. Go back to the original. A press article that talks about a study is not the study. A summary of a summary always distorts things a little. The closer you are to the original source, the better.
5. Look at who and when. Who wrote this text, and when? A 2012 study may no longer say much about a subject that has changed since. And a source that exists is not necessarily a good source: research itself is now overwhelmed by mass-produced texts.
In 2023, in New York, two lawyers filed a document with a court drafted with the help of ChatGPT, which cited several court rulings. Problem: some had never existed. The assistant had invented them, with perfectly credible names and references. The judge imposed a fine on the lawyers. The case made headlines around the world and has since served as a warning in law firms.
The right reflex
You can also ask the assistant directly: "Give me the exact link to this study" or "Quote me the precise passage". If it hesitates, contradicts itself or changes the reference, you have your answer.
And for anything that really matters, your health, your money, an important decision, do not stop at the AI's answer. Use it as a starting point, then go and look at the source yourself.
What to take away
An AI that cites a source has not necessarily verified that source. That work remains yours.
The good news is that it rarely takes more than two minutes. And those two minutes make all the difference between being informed and being convinced.