You ask an AI tool a straightforward question, it gives you a confident, well-written answer — and later you discover the key fact in it was completely made up. The study it cited doesn't exist. The quote was never said. The figure was invented on the spot.
That sinking feeling — I almost used that — is what brings most people to the word "hallucination." It's the single most important AI behaviour to understand, because it's the one most likely to embarrass you in front of a client or a boss.
The good news: once you understand why it happens, it stops being mysterious and starts being manageable. This piece is part of our Terminology Tamer series — the same plain-English treatment we gave large language models and AI agents. By the end, you'll know what a hallucination actually is, why even the best tools do it, and the handful of habits that catch them before they cost you.
The one-sentence answer
An AI hallucination is when an AI tool produces text that is fluent, confident, and plausible — but factually wrong or entirely invented.
The tricky part is in that sentence: fluent, confident, and plausible. A hallucination doesn't look like an error. It looks exactly like a correct answer. There's no flashing red light, no "I'm not sure about this bit." The made-up court case is written in the same calm, authoritative tone as the real one.
That's what makes it dangerous, and it's why "just read the output" isn't enough on its own.
Why even the best tools hallucinate
Here's the part that reframes everything: hallucination isn't a bug that will be patched away next year. It's a side effect of how these tools work.
As we unpacked in the LLM guide, a language model doesn't look facts up in a database. It predicts the most plausible next words, one after another, based on patterns in everything it was trained on. Most of the time, the most plausible continuation also happens to be true — because true statements are well represented in the training text.
But the model has no separate sense of "is this actually true?" running alongside it. It's optimising for plausible, not for correct. When the two line up, you get a right answer. When they diverge — an obscure fact, a specific statistic, a precise citation — the model will still produce the most plausible-sounding text, even if that means inventing a source that fits the pattern of a real one.
Put simply: the tool isn't lying to you. Lying requires knowing the truth and choosing otherwise. It's doing exactly what it was built to do — generating fluent text — and sometimes fluent text isn't factual.
🧠 Quick Challenge: True or false — turning on an AI tool's web-search feature means it can no longer hallucinate.
- A) True
- B) False
Answer: B) False. Web search reduces hallucinations a lot, because the tool can ground its answer in live sources — but it doesn't eliminate them. The model can still misread a source, blend two sources together, or summarise a page inaccurately. As we covered above, it's optimising for plausible text, and that pressure doesn't switch off just because a source is in front of it.
What a hallucination actually looks like
Knowing the shapes they take makes them far easier to spot. These are the ones that catch professionals out most often:
- Invented citations. A study, author, or journal that sounds exactly right and does not exist. Common when you ask for "sources" or "research showing...".
- Fabricated statistics. A precise-looking figure ("73% of managers report...") with no real survey behind it. Specific numbers feel trustworthy, which is what makes these costly.
- Made-up quotes. Words attributed to a real person who never said them.
- Confident wrong details. Wrong dates, wrong job titles, a feature a product doesn't have, a clause that isn't in the contract.
- Plausible-but-fake specifics. Case numbers, page references, API parameters, URLs that 404 — anything precise enough to look verifiable but invented to fit.
Notice the thread: hallucinations cluster around specifics. Vague summaries are usually fine. It's the exact figure, the named source, the precise quote where you need to be on guard.