“Hallucination” is one of the more dramatic words to have entered everyday professional vocabulary in the last few years. It sounds alarming — the AI is hallucinating? — in a way that makes the whole technology feel unreliable.
The reality is more mundane, and more manageable, than the word suggests.
What actually happens
Large language models — the technology behind ChatGPT, Claude, Gemini, and Copilot — work by predicting what words should come next based on patterns learned from vast amounts of text. They’re extraordinarily good at this. They’re so good at it, in fact, that they can produce text that reads as completely plausible even when the specific facts it contains are wrong.
The model isn’t checking its answers against a database of verified facts before responding. It’s generating language that fits the shape of a correct answer. Sometimes that language is accurate. Sometimes it isn’t. And the model delivers both with equal confidence.
What hallucination looks like in practice
It’s rarely dramatic. It usually looks like:
A statistic that’s slightly off — a real figure but misremembered or from the wrong year. A source that almost exists — a paper by a real researcher on a real topic, but the specific paper cited doesn’t actually exist. A quote attributed to someone who never said it, in language that sounds entirely like them. A fact that was true at some point but is no longer accurate.
The common thread: plausible, confident, wrong. Which is more dangerous than obviously wrong, because it gets past your guard.
Should you worry about it?
Not excessively — but you should respect it. The practical rule is simple: verify anything specific before you rely on it.
Dates, statistics, names, quotes, legal or regulatory details, medical information, financial figures — all of these are worth checking independently. For general drafting, summarising, brainstorming, and thinking through problems, hallucination is much less of a concern because the output doesn’t depend on specific factual accuracy.
Think of it this way: if you asked a very capable new colleague who had read a great deal but sometimes misremembered details to help you with something, you’d check their specific claims before putting them in a report. That’s exactly the right relationship to have with AI.
The one thing that reduces the risk most
Giving the AI the source material to work from, rather than asking it to recall facts from memory. If you paste in a document and ask AI to summarise it, the hallucination risk drops significantly — it’s working from what you’ve given it rather than from what it thinks it knows.
For more on AI terminology in plain English, the AI in Plain English glossary covers the terms that come up most. And if you’re still getting started with prompts, What Is a Prompt? is the right place to begin.