Why AI makes things up, and how to make it stop
AI hallucinations explained in plain English: why models invent facts and sources, and five prompt habits that make it much less likely.
, 2 min read
Ask an AI model for a source and you may get a perfect-looking citation for a paper that does not exist. This is called a hallucination, and it is the single biggest reason people stop trusting AI answers. It helps to know why it happens.
Models predict, they do not look up
A language model writes by predicting the next most likely word, over and over, based on patterns in its training data. It is not searching a database of facts. Most of the time, the most likely next word is also the true one. But when the model does not know something, it does not stop. It keeps predicting, and produces something that looks like a true answer.
A fake citation is a good example. The model has seen thousands of real citations. It knows exactly what one looks like: authors, a year, a journal, a title. So it writes one. The shape is right. The content is invented.
When it happens most
- Specific details: exact numbers, dates, quotes, names, page numbers and links.
- Rare topics that appear little in training data.
- Recent events after the model's knowledge cutoff.
- Leading questions, where you have signalled the answer you want.
Five habits that help
1. Give it the source
The most effective fix by far. Paste the document, page or data, and say "answer only from this; if it is not there, say so". This is called grounding.
2. Give it permission not to know
Add "If you are not sure, say you are not sure." Models are trained to be helpful, and a direct invitation to admit uncertainty makes them much more likely to do so.
3. Ask for quotes
"Support each point with a direct quote" makes invention much harder, and makes checking much easier. The summarise with quotes prompt is built for this.
4. Turn on search when it matters
Many assistants can search the web or your files. For anything recent or specific, use it, and check the links it gives you open to the right page.
5. Check the specifics
Treat names, numbers and citations as unconfirmed until you have checked them. The fact-check claims prompt pulls them out into a list so you can check them quickly. For a full routine, see how to check an AI answer before you rely on it.
It is getting better, not gone
Newer models make things up less often, and many now say when they are unsure. But no model is immune. The habits above take seconds, and they turn AI from something you have to double-check from scratch into something you can trust with a quick look.