An AI hallucination is when a chatbot gives an answer that sounds confident but is false: an invented statistic, a fake quote, a book that does not exist or a court case that never happened. It happens because language models generate the most plausible next words, not verified facts. You cannot switch hallucinations off, but you can make them rarer and easier to catch.
Why AI hallucinates
A large language model learns statistical patterns of language. When you ask a question, it predicts what a good answer would look like. If the right information was rare in its training data, is newer than its knowledge cutoff, or is missing from the context you gave it, the model still produces fluent text, and fluent text can be wrong. Research from OpenAI has also pointed out that training and evaluation often reward guessing over admitting uncertainty, which pushes models to answer anyway.
Common types of hallucination
| Type | Example |
|---|---|
| Invented facts | A wrong date, figure or name stated as fact |
| Fake sources | Citations, URLs or studies that do not exist |
| Misread documents | A summary that contradicts the file you uploaded |
| False reasoning | A calculation or logic chain with a hidden error |
| Overconfident gaps | A detailed answer about an obscure topic the model barely knows |
Real consequences
The best-known example came in 2023, when lawyers in a US federal case filed a brief citing court decisions that ChatGPT had invented; the judge sanctioned them. Similar problems have appeared in academic papers with fake references and in customer-service bots giving customers wrong policy information. The pattern is the same: an AI answer was used without checking.
How to reduce hallucinations
- Give the model the source. Paste the text or upload the document and tell it to answer only from that.
- Turn on web search for anything recent, and ask for links.
- Allow uncertainty. Add “If you are not sure, say so” to your prompt.
- Ask for quotes. “Quote the exact sentence that supports this” makes invented details easier to spot.
- Break big questions into smaller ones so each step can be checked.
- Use a reasoning model for math and logic, or ask it to use a code tool for calculations.
Our guide on how to write prompts covers these techniques in more depth.
How to catch hallucinations
Check names, numbers, dates and citations against the original source, especially in anything you will publish or act on. Click every link. Ask a second model the same question; disagreement is a warning sign. An all-in-one AI platform makes that cross-check quick. Be most careful with obscure topics, recent events, legal, medical and financial details, and anything that sounds surprisingly convenient.
When small errors matter less
For brainstorming, drafting or rewriting your own text, an occasional error costs little because you review the output anyway. The risk grows when AI output goes straight to customers, into code that runs in production, or into decisions with legal or financial weight. Match the amount of checking to the cost of being wrong.
To understand where these errors come from in a specific product, read how ChatGPT works.
Frequently asked questions
What is an AI hallucination?
An AI hallucination is an answer that sounds confident and plausible but is false or unsupported, such as an invented quote, statistic, citation or legal case.
Can AI hallucinations be completely prevented?
No. They can be reduced with web search, grounding in your own documents, clear prompts and verification, but every current language model can still produce false statements.
Which AI hallucinates the least?
Rates change with every model release and depend on the task. Tools that ground answers in sources, such as Perplexity or a chatbot with search turned on, make errors easier to spot, but they do not eliminate them.