The single most important thing to understand about AI: it can be confidently wrong. Knowing why — and how to catch it — is the difference between AI helping you and embarrassing you.
What is a hallucination?
A hallucination is when an AI states something false as if it were fact — a made-up statistic, a fake citation, a non-existent feature, a plausible-sounding wrong answer. It's not lying; it doesn't know it's wrong.
Why it happens
Language models don't "look things up." They generate the most likely next words based on patterns learned in training. Usually that produces something true — but when the model is unsure, it fills the gap with something that sounds right rather than admitting it doesn't know. There's no built-in fact-checker.
Where hallucinations are most dangerous
- Facts, dates and statistics — especially specific numbers.
- Citations and quotes — models invent realistic-looking sources.
- Legal, medical and financial advice — high stakes, high risk.
- Code — it may call functions or libraries that don't exist.
- Anything recent — beyond the model's knowledge cutoff.
How to catch them
- Verify anything that matters — treat facts as claims to check, not truths.
- Ask for sources — then actually open them; invented links won't resolve.
- Use grounded tools for research — Perplexity cites live sources, which makes checking easy.
- Cross-check across tools — ask ChatGPT and Claude the same thing; disagreement is a red flag.
- Be specific — vague prompts invite vague, made-up answers.
- Watch for over-confidence on the obscure — the more niche the topic, the higher the risk.
The right mindset
Treat AI as a brilliant, fast intern: great for drafts, ideas and first passes — but everything ships past your judgment. AI writes; you verify. Get that habit right and hallucinations go from a liability to a non-issue.
New to this? Start with our plain-English guide to AI.