What it means
A hallucination is when an AI system presents something false or invented as if it were plain fact.
The classic examples are small and specific. A chatbot invents a book and credits it to a real author. It quotes a study that was never published. It supplies a statistic or a web link that leads nowhere. You’d say it hallucinated the book, the study or the link. It sounds just as calm as when it’s right, which is what makes it awkward.
The cause lies in how large language models work. They’re built to produce the text most likely to come next. They have no separate fact checker. When the model has seen plenty about a topic, likely text and true text usually line up. When it hasn’t, it still produces something likely, and likely isn’t the same as true.
Think of a student who hasn’t read the set book but has to write the essay anyway. They know how essays about novels sound, so they write a fluent one, with quotes that feel right. The quotes aren’t in the book. The student isn’t quite lying, just filling a gap with something plausible and hoping nobody checks.
Where it came from
The everyday word means seeing or hearing things that aren’t there. An earlier technical use came from computer vision research, where “hallucinating” detail meant filling in plausible features, such as sharpening a blurry photo of a face. There, it was often the goal rather than a fault.
By the late 2010s, researchers working on machine translation and automatic summaries were using it critically. A system hallucinated when its output included content the source didn’t support, like a summary quoting a figure the original article never gave.
The usual account is that the word reached everyone else in early 2023, once ChatGPT had millions of users and screenshots of confident nonsense were everywhere. News stories followed, including reports of lawyers who filed court documents citing cases that did not exist. Cambridge Dictionary chose “hallucinate” as its Word of the Year for 2023 because of this AI sense.
Not everyone likes the term. Some researchers argue it makes software sound too human, since hallucinating implies a mind having experiences. Alternatives such as “confabulation” have been suggested, but “hallucination” has stuck.
Here’s how it compares with nearby words.
| Word | What’s happening | Intent to deceive? |
|---|---|---|
| Hallucination | An AI produces false content that sounds plausible | No |
| Lie | Someone says something they know is false | Yes |
| Mistake | An ordinary error, often easy to spot | No |
| Confabulation | Filling a memory gap with invented detail, unaware it’s invented | No |
| Misinformation | False information spreading, whatever its source | Not necessarily |
How people actually use it
- “It hallucinated half the references.” Work chat, after someone checked the citations.
- “Is this real or did it hallucinate it?” Now a reflex with any surprising AI answer.
- “Grounded in your documents to reduce hallucinations.” Product page wording, usually meaning the tool looks things up first, a method called RAG.
- “Answers may contain hallucinations. Verify before relying on them.” A typical disclaimer under an AI tool.
- “Chatbot invented legal cases, court hears.” A headline format that became familiar from 2023.
- “Ben just hallucinated a deadline that doesn’t exist.” A joke aimed at a human colleague.
In a sentence
Nadia: the summary says the report recommends a four-day week
Chris: it doesn’t. I wrote it
Nadia: so the AI hallucinated that
Chris: with great confidence, yes
Tom: it gave me a link to the paper
Ade: does the link work
Tom: no
Ade: there’s your hallucination
Ellie: can I use it for my essay
Raj: for ideas, sure. for quotes and dates, check every single one
Common misconceptions
- It isn’t a rare glitch. Hallucination comes from the basic way these models generate text. Newer models do it less, but none is free of it.
- The AI isn’t lying to you. Lying needs an intention to deceive, and the model has none. It produces plausible text, and sometimes plausible is wrong.
- A citation proves nothing on its own. Models can invent sources, page numbers and links as easily as facts. Check that the source exists and says what’s claimed.
- “Don’t make things up” isn’t a fix. Careful prompt engineering helps a little, and so do tools that let the model search or read your files. None of them removes the risk.
Related terms
- Confabulation: a psychology term for filling memory gaps with invented detail, proposed as a more accurate label.
- Grounding: tying an AI’s answers to specific, trusted sources.
- RAG: retrieval-augmented generation, where a model looks up documents before answering.
- Sycophancy: when an AI tells you what it thinks you want to hear, even agreeing with false claims.
- AI slop: low-effort AI content churned out in bulk, often unchecked and full of errors.
Questions people ask
Why do AI chatbots hallucinate?
AI chatbots hallucinate because they generate plausible-sounding text rather than checking facts against a reliable source. When they lack solid information, they still produce an answer, and it can be wrong.
Can you stop an AI from hallucinating?
AI hallucinations can be reduced but not eliminated. Giving the model trusted documents to work from, asking for sources and checking key facts yourself all help.
Is an AI hallucination the same as lying?
An AI hallucination isn’t a lie in the usual sense. Lying means knowingly saying something false, and an AI model has no intention to deceive.
Do newer AI models still hallucinate?
Newer AI models hallucinate less often than early chatbots, especially when they can search the web or read documents you supply. They still do it, particularly with obscure topics, exact numbers and citations.
Why was “hallucinate” Cambridge Dictionary’s Word of the Year?
Cambridge Dictionary picked “hallucinate” as its Word of the Year for 2023 because of its newer AI meaning: a system producing false information.
The short version
A hallucination is when an AI confidently gives you something false, from an invented quote to a source that doesn’t exist. It happens because language models predict plausible text rather than checking facts. Treat AI answers as a first draft and verify anything that matters.