How to Spot Hallucinations and Use AI More Confidently.
Last month we wrote about choosing your first real AI project at work – picking something small, low-risk, and seeing what these tools can actually do. One of the things we mentioned in passing was that AI tools are confident even when they’re wrong. That’s worth a whole article of its own, because it’s the single biggest reason people lose trust in AI after a promising start.
The technical term is hallucination – when an AI tool produces something that sounds completely plausible but is, on closer inspection, made up. A fake statistic. A book that doesn’t exist. A legal case that was never decided. The output reads like the real thing because the AI is genuinely trying to be helpful – it just doesn’t have a reliable way to know when it’s at the limit of what it actually knows.
Why It Happens
To understand hallucinations, it helps to know what AI chatbots are actually doing. They aren’t looking things up in a database – they’re predicting the most likely next words in a sentence, based on patterns from the enormous amount of text they were trained on. Most of the time, the most likely next words are also true. But when the AI hits a question where it doesn’t have solid information, it doesn’t say “I don’t know.” It produces the most plausible-sounding answer instead.
Stanford’s annual AI Index tracks how often this happens and how it varies by model, and the trend over the past two years has been towards fewer hallucinations as the technology improves – but they haven’t gone away, and probably won’t anytime soon.
The Patterns to Watch For
Hallucinations cluster in a few common areas:
Specific dates, statistics, and citations. If an AI quotes a number, a percentage, or a study, treat it as a hypothesis to verify rather than a fact. The number sounds precise; that’s exactly what makes it dangerous.
Quotes from real people. AI tools will happily generate quotes from named individuals that those people never said. High-profile cases have included lawyers being sanctioned for filing legal documents containing entirely fabricated case citations the AI produced.
Recent events. Most AI tools have a knowledge cut-off date and limited or no access to current events. Ask about something from last week and you may get a confidently wrong answer rather than a “sorry, I don’t know.”
Highly specific technical detail. The deeper into a specialist topic you go, the more likely you are to get an answer that sounds expert but isn’t.
Simple Habits That Catch Most Errors
You don’t need to become a fact-checker. A handful of habits will catch most hallucinations before they cause problems.
Ask for sources. “Where did that statistic come from?” or “Can you cite the source?” If the AI can’t produce a real, checkable source, treat the claim as unverified. If it does cite a source, spend 30 seconds checking the source actually exists and says what the AI claims.
Verify anything you’d quote externally. The rule of thumb: if you’d be embarrassed to find out it was wrong, check it independently before using it.
Push back when something seems off. AI tools often correct themselves when you ask “are you sure about that?” – which tells you something about the confidence level on the first answer.
Pro Tip: Try asking the AI for sources, then deliberately mention a source you’ve made up. A good response is one where the AI says it isn’t familiar with that source. A poor response is one where the AI confidently summarises a paper that doesn’t exist. It’s a useful quick test of how much to trust an answer.
Confident, Not Cynical
None of this means AI tools aren’t useful. They’re genuinely valuable – drafting, summarising, brainstorming, explaining tricky concepts. But they’re useful in the way a very knowledgeable colleague who occasionally gets things wrong is useful: brilliant for first drafts, hopeless if you take their word for everything. Keep your judgement switched on, and you’ll get the best out of them without the risk of getting caught out.
If you tried last month’s “first AI project” suggestion and ended up impressed, this is the natural next step: not less use, but more careful use. And that’s where the real long-term value of these tools lives.