Confident, fluent, and wrong
Sometimes a chatbot states something false as if it were plain fact. It might name a book that was never written, give a date that's off by ten years, or describe a feature a product doesn't have. People call this hallucination. The word is a little odd, since the model isn't seeing things. It just means the model made something up.
What makes hallucinations tricky is that they look exactly like correct answers. The tone is just as calm. The grammar is just as clean. There's no warning label.
Why it happens
Recall how a language model works: it predicts a likely next piece of text, one piece at a time. Its goal is text that fits, not text that has been checked. Usually the most likely answer is also the true one, because true things get written about often. But when the model has no solid pattern to lean on, it still produces something, and that something is whatever sounds most plausible.
Say you ask for a study on a narrow topic. The model knows what study titles look like. It can build one that sounds perfect, with a real-sounding author and journal, even if no such study exists. In that moment, plausible wins over true.
Training can add to this. A model shaped to be helpful may lean toward giving some answer rather than saying "I don't know." Newer models are often better at admitting doubt, but none are free of the problem.
Common forms
Hallucinations often show up as:
- Fake sources. Citations, links, quotes, or court cases that don't exist, or real ones credited to the wrong person.
- Wrong numbers. Dates, prices, or statistics that are close but off, or invented outright.
- Invented details. Extra facts added to fill gaps, like a made-up middle name or a step in a process that isn't real.
Warning signs
You can't always spot a hallucination just by reading. But some things should make you look twice:
- Very specific details on an obscure topic.
- A source, quote, or link you can't find anywhere else.
- Answers that change when you ask again.
- Anything you plan to act on, such as health, money, or legal matters.
When the stakes are real, check the claim against a source you trust.
Try it
Ask a chatbot for three books about a narrow topic you know well, with authors and years. Then search for each one. Count how many are real, and notice whether any details are slightly off.