How to fact-check AI answers.

AI assistants are genuinely useful — for drafting, explaining, summarising and thinking out loud. But they are not a source of truth. A chatbot doesn't look things up in a database of verified facts; it predicts the most plausible next words based on patterns it learned from text. Most of the time that lands on something correct, but the same machinery will happily produce a fluent, confident, and completely made-up answer — an invented statistic, a fake quote, a citation that doesn't exist. This guide explains why that happens, how to spot it, and a simple routine for checking anything that actually matters.

What is an AI "hallucination"?

A hallucination is when an AI states something that sounds plausible but is false or fabricated. The word makes it sound dramatic, but the mechanism is mundane: it's not the model being sneaky, and it isn't a rare glitch either. It genuinely has no built-in sense of true versus untrue. It's optimising for text that reads right, and confident, well-formed prose reads right whether or not the underlying facts are real.

In practice hallucinations show up as: made-up academic citations or court cases (with realistic-looking authors, titles and dates); wrong dates and timelines; invented statistics stated to a suspiciously precise decimal; fake product features or pricing; misattributed quotes; and biographies that blend two different people together. The giveaway is that all of it is delivered in the same calm, authoritative tone as the parts that happen to be correct.

When AI is most likely to be wrong

You don't need to distrust everything equally. Hallucinations cluster around a few predictable situations. Raise your guard when a question involves:

• Very recent events — anything after the model's training cutoff, unless it's actively searching the web.
• Niche or specialist topics where there's little text to have learned from.
• Specific numbers, names, dates and quotes — the exact details are the easiest things to get subtly wrong.
• Legal, medical and financial specifics, which change over time and vary by country.
• "Facts" offered with no source, especially ones that are oddly convenient for the answer.
• Anything it seems too eager to answer — a good sign a model is guessing is that it never says "I'm not sure."

How to spot a hallucination

There's no perfect detector, and you shouldn't try to eyeball your way to certainty — verification is what settles it. But a few tells come up again and again, and they're useful for deciding what deserves a closer look:

• Confidence without sources. A strong, specific claim and nothing to back it up is the classic pattern.
• Oddly specific detail. Precise figures, exact quotes or full citations that appear from nowhere deserve a second look — precision is easy to fake.
• Answers that don't quite fit the question. If the reply drifts slightly off what you asked, the model may be pattern-matching rather than reasoning.
• Citations that don't resolve. Links that 404, DOIs that lead nowhere, or a real source that turns out to say something different are the single clearest signal.

A simple fact-checking routine

For anything that matters, run the answer through a quick checklist rather than taking it at face value:

• Ask for sources — then check they exist. Request a citation for each claim, then actually open it and confirm it says what the AI says it says.
• Verify the specifics against a primary source. Names, dates and numbers should be traced back to the original document, official site or dataset, not another summary.
• Cross-check in a second tool. Paste the same question into another assistant and compare. For research that needs citations, a tool like Perplexity that answers with linked sources makes checking faster.
• Make it argue against itself. Ask "what might be wrong with that answer?" or "list what you're least sure about." Models will often flag their own shaky spots when invited to.
• Never paste unverified facts into anything important. A report, an email to your boss, a legal or medical decision — verify first, publish second.

Prompts that reduce hallucinations

You can also cut down on invented answers up front with a few small instructions:

• Give it permission to say "I don't know." Add: If you're not sure, say so instead of guessing.
• Ground it in your own material. Paste the text and add: Only use the information I provide above — don't add outside facts.
• Require a source per claim. Ask: Cite a source for every factual statement, and flag anything you can't source.

None of this makes AI untrustworthy — it makes it a fast, capable assistant you supervise rather than an oracle you obey. Build the checking habit and you get the speed without the surprises. From here, it's worth reading our guide on how to write better AI prompts and, for cited research, our Perplexity prompts.