Short answer
An AI hallucination is when an AI states something false as if it were fact. It can invent a book title, a court case that never happened, or a statistic, and attach a believable source. This happens because a chatbot is not a lookup tool. It is a program that writes the most plausible next words. Newer models do it less often, but none have stopped, so check every number and source yourself.
Why does it happen?
Chatbots like ChatGPT are built from a language model that read a huge amount of text and learned which words tend to follow which. When you ask a question, it does not open an encyclopedia and look up the answer. It builds a reply one word at a time, following the patterns it learned.
That is why the sentences come out smooth even when the model does not know the answer. Ask about "a 2019 paper by Professor X" and it will write something shaped like a paper title, whether or not that paper exists. It is a bit like a person who bluffs instead of saying "I don't know." The difference is that the AI has no intent to deceive.
In a September 2025 paper, OpenAI pointed to how models are graded as one cause. On a test, "I don't know" scores zero and a lucky guess sometimes scores a point. If an AI is judged only on accuracy, it learns to guess when it is unsure.
Questions where it goes wrong most
- Requests for sources: papers, books, article links, case numbers. A real-looking citation that does not exist is the most common failure.
- Exact numbers: statistics, dates, prices, populations, tax rates.
- Recent events: news, prices or law changes after the model's training cutoff. Without web search turned on, it cannot know.
- Lesser-known people, companies and local businesses: the less material there is, the more likely it is to blend similar things together.
- Long calculations: with many steps, one early mistake carries through and the final answer still sounds sure.
Try it yourself
Ask about something that almost certainly does not exist and watch the reply. Modern chatbots often say they cannot confirm it, but a slightly different wording can still get a made-up answer.
Who wrote the 2018 Stanford paper on the link between cat whisker length and intelligence, and what did it conclude?
If the AI gives author names and a conclusion without being able to confirm the paper exists, that is a hallucination.
How to ask so it happens less
You cannot prevent it completely, but how you ask makes a big difference.
- End your question with "If you don't know, say you don't know."
- Turn on web search. ChatGPT has search, Claude has web search, and Gemini uses Google Search by default. When the answer comes with source links, open them and check that the page says what the AI says it does.
- Give it the material. Paste the PDF or article text and say "Answer only from this text." That leaves much less room to invent.
- Ask one thing at a time. When you mix questions, a mistake in the first answer carries into the next.
- For math, ask it to show the steps, or recheck with a spreadsheet or calculator.
From your last answer, list only the claims that need fact-checking (numbers, dates, names, sources) in a table. Mark anything you are not sure about as "needs verification."
How to check what it told you
If a link does not open, or it opens and the claim is not on the page, treat that part as made up.
| What the AI gave you | Where to check |
|---|---|
| Papers, books | Google Scholar (scholar.google.com), WorldCat (worldcat.org), or a library catalog |
| Laws, court cases | The court's own records or an official legal database, not a summary from the AI |
| Statistics | The agency or organization that publishes the data, usually on a .gov or .org site |
| Taxes, benefits | The official government site for your country, or the agency that runs the program |
| Health, medication | A doctor or pharmacist, and the official drug regulator's site |
| Articles, links | Open the link. Does it load, and does the page say what the AI said? |
When it is still fine to use
Hallucination matters when facts matter. For brainstorming, polishing a sentence, drafting an email or summarizing, where you can judge the result right away, it is much less of a problem. Even when you give the AI a document to summarize, it sometimes adds something that is not in the original, so compare the two when the document is important.
Related terms
- LLM (large language model)
- The program inside ChatGPT, Claude and Gemini that actually writes the text. It learns from a lot of writing to predict the next words.
- Training cutoff
- The date of the last material the AI studied. Without web search, it does not know what happened after that.
- Grounding
- Making the AI answer from real material, like search results or a document you gave it. It is the main way to reduce hallucination.
FAQ
Do paid versions hallucinate less?
They can, but they still do it. Better models more often say "I don't know," but numbers and sources still need checking.
If the AI gives me a link, can I trust it?
Not until you open it. Links from a search-enabled answer usually exist, but the page may say something different from the AI's summary.
Why is it called "hallucination"?
The word borrows from people seeing things that are not there. For AI it just means a confident, plausible-sounding wrong answer. "Making things up" is a fair way to think of it.
Can I paste AI answers straight into homework or a report?
Not without checking the facts. You could end up citing sources that do not exist. In the US and elsewhere, lawyers have been sanctioned for filing court papers that cited cases an AI invented.
Checked against sources on 2026-10-30. Sources: OpenAI, Why language models hallucinate (2025-09), Kalai et al., Why Language Models Hallucinate (arXiv 2509.04664). How we write