AI hallucinations: why your chatbot lies and how to catch it
An AI hallucination is when a chatbot states something false with complete confidence: a made-up statistic, a fake source, a quote that was never said, a feature a product doesn't have. It happens because these tools are built to produce plausible text, not verified facts. You catch it by asking for sources, checking anything specific, and treating fluent confidence as a warning sign rather than proof.
I'm Rafaela Buena, I teach AI to non-technical people in Amsterdam, and hallucinations are the single topic where I see the most fear and the least understanding. Once you know why they happen, they get much easier to spot.
AI hallucinations explained: why it happens
A language model doesn't look things up. It predicts what text should come next based on patterns from everything it's read. Most of the time, the most plausible next text is also true. Sometimes it isn't, and the model has no internal alarm that goes off when it crosses that line.
Ask it for a study supporting your point and it will produce something that looks exactly like a study: authors, year, journal, plausible title. Whether that study exists is a separate question the model never asked itself.
The fluency is the trap. Wrong answers arrive in the same confident tone as right ones.
Where hallucinations show up most
Specific numbers and statistics. "73% of small businesses..." Check it.
Citations and sources. Papers, articles, books, URLs. The most common hallucination of all.
Quotes attributed to real people. Especially ones that sound a bit too perfect.
Niche or recent facts. The less the model has read about something, the more it improvises.
Details about your own business. It doesn't know your pricing, your policies, or what happened last quarter unless you tell it.
Anything you asked it to be certain about. "Are you sure?" often produces a more confident version of the same mistake.
Five ways to catch them
Ask for sources, then check one. If the source doesn't exist or doesn't say what it claims, discard the whole answer and start again.
Give it the material. Paste in the document, the data, the page. Hallucinations drop sharply when the model is working from something real instead of from memory.
Ask it to flag uncertainty. "Tell me which parts of this you're confident about and which you're guessing." It's imperfect, but it catches a lot.
Never publish a specific claim you haven't verified. Numbers, names, dates, quotes. This is the rule I drill in every 1-1 session, and it's less about distrusting the tool than about keeping your own name clean.
Use search-connected tools for facts. Modern chatbots can search the web. When you need something current or checkable, turn that on and read what it cites.
What hallucinations don't mean
They don't mean the tool is useless. For drafting, summarising material you provide, structuring your thinking, and rewriting, hallucination risk is low, because the model is working with your input rather than inventing facts.
They don't mean you need to check every word. They mean you need to check every claim. There's a difference, and learning it is what separates people who use AI confidently from people who either trust it blindly or fear it entirely.
The mindset that works
Treat the chatbot like a brilliant, fast colleague who sometimes bluffs. You wouldn't fire them. You'd learn where they bluff and check there. Same here.
FAQ
Do all AI chatbots hallucinate?
Yes. Claude, ChatGPT, Gemini, all of them. Rates vary by model and task, and they've improved a lot, but no current tool is immune. Design your habits around that.
Can I stop AI hallucinations completely?
Not completely. You can reduce them dramatically by providing source material, asking for sources, and using search-enabled modes for factual questions. The rest is verification.
Is it safe to use AI for research?
Yes, as a starting point and a structuring tool. Not as a final source. Use it to find angles and draft questions, then verify anything you plan to rely on or repeat.
If hallucinations are the reason you don't trust AI with real work, bring that worry to a 1-1 session. An hour on your own tasks, with a checking habit built in, and the fear tends to turn into a shrug.
