If you use AI for anything real, you will run into a hallucination. It's the single most common way AI lets people down. This lesson explains what one is, why it happens, and how to catch it before it ends up in your work.
What a hallucination actually is
A hallucination is when an LLM gives you a confident answer that is wrong.
Not a lazy answer. Not a cautious one. A confident, detailed, fluent answer that sounds right and isn't.
Common examples:
- Citing a book or paper that doesn't exist.
- Quoting a person who never said that thing.
- Inventing a software feature that isn't in the product.
- Making up a statistic ("73% of small businesses…") that was never measured.
- Confidently giving you the wrong year, date, or price.
The giveaway is usually plausibility with too much detail. It reads like the AI has researched this thoroughly. It hasn't.
Why hallucinations happen
Remember the previous lesson: an LLM doesn't have a database. It predicts the next token. If it's seen lots of similar-looking text during training, its prediction will usually be right. If it hasn't, it will still make a prediction. It doesn't have a "don't know" button.
Imagine a bright student sitting an exam they haven't revised for. They're clever, they've read a lot, and they'd rather guess confidently than hand in a blank page. Sometimes they guess right. Sometimes they guess with conviction and are completely wrong. An LLM behaves the same way.
Three situations make hallucinations more likely:
- Very specific facts. Names, dates, numbers, citations, prices. Anything where "roughly right" is still wrong.
- Recent events. The AI's training data has a cut-off date. Ask it about anything after that and you're rolling the dice.
- Niche topics. The less training data there was on a subject, the more creative the AI will get.
Hallucinations are not a minor quirk. They're the central thing to understand about AI.
How to spot one
Four cheap tests you can run on any AI answer.
- Does it cite a source? If yes, click through. If the source doesn't exist, that's a hallucination.
- Is the number suspiciously round? "Exactly 50%", "doubled in five years", "£1,000 per month on average". If you can't find that exact number in a real report, be sceptical.
- Would you have heard of this? If the AI mentions a "famous" person, book, or study you've never encountered in the topic you know, that's a flag.
- Ask again, in a new chat. Hallucinations are often inconsistent. Ask the same question fresh and see if the answer stays the same.
How to reduce them
You can't eliminate hallucinations. You can make them rarer.
- Give the AI the source material. Paste the article, upload the document, link the web page. The AI can't hallucinate facts that are sitting in front of it.
- Ask for a quote, not a paraphrase. "Quote the exact sentence from the document that supports this" forces the AI to point to real text.
- Use a model with web search turned on for anything time-sensitive. Gemini and ChatGPT both support this.
- Lower the stakes of the mistake. Use AI for brainstorming and first drafts, where a wrong fact is easy to catch. Don't use it unchecked for legal, medical, or financial advice.
The one rule
If the cost of being wrong is high, you always verify. Every fact, every number, every quote.
The next lesson gives you a 5-step checklist for doing exactly that.