Typical examples are a citation that doesn't exist, a date that's a year out, or a clause the model says is in a contract when it isn't.
Why do AI models hallucinate?
A large language model generates the most plausible continuation of the text in front of it. Most of the time, plausible and correct line up. When they diverge, you get a sentence that reads perfectly and is wrong. Ask about a commercial lease and the model may describe a break clause, because leases often have one, whether or not yours does.
The research term comes from work on summarisation, where models were observed producing content the source didn't support, for example in Maynez et al., 2020.
There's no change in tone when the model is wrong, so you can't tell from the writing alone which sentences to doubt.
How should it change the way you use AI?
Treat facts, figures, names, dates, legal references and quotations as unverified until you've checked them. Use structure, tone and phrasing straight away, since that's where AI saves you time with little risk.
An example
In 2023, two New York lawyers filed a brief in Mata v. Avianca that cited court cases ChatGPT had invented, complete with fabricated quotations. The judge in the Southern District of New York fined the lawyers and their firm $5,000. Each fake case looked like a real one, because case citations follow a pattern the model could imitate.
The everyday version is smaller. You ask for a summary of a regulation, and it cites a section number that sounds right. The summary may be broadly accurate while the section number is made up. Checking takes one search on the official legislation site; quoting it unchecked in a board paper can take weeks to unwind.
Is a hallucination the same as a lie?
No. A lie requires knowing the truth and choosing to contradict it, while a hallucination is a plausible guess with no intent behind it. That changes the fix. Telling the model to stop making things up does little. Giving it source material through retrieval-augmented generation, asking it to say when it's unsure, and checking what matters all help.
Common questions
Can hallucinations be eliminated?
Not with current methods. Grounding the model in supplied documents and requiring citations reduces them, and guardrails can catch some, but the risk doesn't reach zero.
Where do hallucinations usually appear?
In specifics: numbers, dates, names, quotations, case references and web addresses. Broad general statements are less likely to be wrong in a way that matters to you.
Do bigger models hallucinate less?
Generally they make fewer obvious errors, but the problem remains. A wrong answer from a strong model can be harder to doubt, which is why a human check on anything that matters stays useful.