Skip to content
AI Digested
GuideBeginner3 min read

Why AI makes things up (hallucinations) and how to catch it

AI chatbots sometimes invent facts, quotes, sources and statistics with total confidence. Here's why it happens, when it's most likely, and simple habits to protect yourself.

First published · AI Digested

Ask an AI chatbot for a list of books on a topic and you might get a perfect reading list, with one or two titles that don't exist, complete with plausible authors and publication years. This is called a hallucination, and it's the single most important thing to understand about AI before you rely on it.

What a hallucination looks like

  • A quote that sounds right but the person never said.
  • A statistic with a precise-sounding number that came from nowhere.
  • A research paper, book or web link that doesn't exist.
  • A historical date that's off by a few years.
  • A made-up feature in a product's instructions.
  • A legal case that was never decided.

That last one really happened. In 2023, two New York lawyers were fined after submitting a court filing containing cases ChatGPT had invented. Similar incidents have happened in courts around the world since.

Why it happens

A large language model isn't a database. It doesn't store facts in neat rows and look them up. It learned patterns from an enormous amount of text and generates answers by predicting, one piece at a time, what text is likely to come next.

Most of the time, the most likely text is also the true text. "The capital of France is…" has only one sensible continuation. But when the model hasn't seen enough about something, or the question is very specific, it still produces fluent, plausible-sounding text, because that's what it was trained to do.

There are a few other reasons:

  • Training rewards confident answers. Research by OpenAI and others suggests models learn that guessing scores better than saying "I don't know".
  • Out-of-date knowledge. A model only knows what was in its training data, up to a cut-off date.
  • Leading questions. Ask "Why did Einstein win two Nobel Prizes?" and some models will invent a reason, rather than tell you he won one.
  • Long, complicated tasks give more room for small errors to creep in.

When to be most careful

Hallucinations are most likely, and most harmful, with:

  • Specific facts: names, numbers, dates, prices, quotes and references.
  • Niche topics where little has been written.
  • Recent events, unless the AI searched the web.
  • Medical, legal and financial questions, where mistakes have real consequences.
  • Anything you'll publish, submit or act on.

AI is far more trustworthy for tasks where you can see whether the result is right: rewording your own text, summarising a document you've given it, brainstorming, or explaining a concept you can cross-check.

Seven habits that protect you

  1. Give it the source material. Paste in the document or article and ask it to answer only from that. Hallucinations drop sharply.
  2. Use search-connected modes. Most chatbots can search the web and show links. Click the links.
  3. Ask for sources, then check them. A real link you open is evidence. A reference you haven't checked isn't.
  4. Tell it that "I don't know" is fine. Adding "If you're not sure, say so" to your prompt helps.
  5. Don't lead the witness. Ask "Did X happen?" rather than "Why did X happen?"
  6. Cross-check important answers with a trusted source, or with a second AI.
  7. Keep humans responsible. For health, legal or money decisions, use AI to prepare your questions, then ask a professional.

Is it getting better?

Yes. Modern models hallucinate much less than early ChatGPT did, especially when they can search, cite sources and reason through problems step by step. Companies measure hallucination rates and publish them in their model reports.

But no current model is hallucination-free, and because the answers have become better, the remaining mistakes are harder to spot. The habit of checking is more important, not less.

Keep going

Common questions

Why is it called a hallucination?

The word caught on because the AI 'sees' something that isn't there. Some researchers prefer 'confabulation' or simply 'making things up', because the AI isn't perceiving anything.

Do newer AI models still hallucinate?

Yes, though much less than early versions, especially when they can search the web and cite sources. No current model is free of hallucinations.

Can I just ask the AI if it's sure?

It can help, as the model may catch its own mistake, but it isn't reliable. Models can apologise and change a correct answer, or double down on a wrong one. Check an independent source.

Sources and further reading

  1. Mata v. Avianca: judge sanctions lawyers for citing fake cases produced by ChatGPT (2023)
  2. OpenAI (2025). Why language models hallucinate

Spotted a mistake or something out of date? Tell us and we will fix it. Read how we research and write our articles.