What a hallucination actually is
A "hallucination" is when an AI model generates information that's false, invented, or unsupported — a fake citation, a wrong date, a function that doesn't exist in a programming library — but presents it with the same fluent confidence as something true.
Why it happens
Language models are trained to predict statistically likely text, not to verify facts against a database. When a model doesn't actually "know" an answer, it doesn't have a built-in way to say "I'm not sure" by default — it generates the most plausible-sounding continuation of the text, which can be wrong while still reading perfectly naturally.
This is different from a person lying, which implies intent. The model isn't trying to deceive you; it's doing exactly what it was built to do — produce likely text — in a case where "likely-sounding" and "true" don't line up.
Where hallucinations show up most
- Citations and sources. Fake paper titles, wrong authors, or URLs that don't resolve are especially common — the model knows what a citation looks like without necessarily having the real one memorized.
- Specific numbers and dates. Precise statistics or dates are easy to state confidently and easy to get wrong.
- Niche or recent topics. The less represented something was in training data, the more likely the model is to fill gaps with plausible-sounding guesses.
- Code that "should" exist. Functions or library methods that sound like they ought to exist, but don't.
How to catch it
- Ask for sources, then check them. Don't just accept a citation — click through or search for it independently.
- Be more skeptical of specifics than generalities. A broad explanation of a concept is usually more reliable than an exact number, date, or quote.
- Cross-check anything decision-critical. If a fact will affect a real decision — medical, legal, financial, or otherwise — verify it through a source you trust independently of the AI.
- Notice unusual confidence on unusual questions. If a topic is obscure and the model answers instantly and precisely with no hedging, that's worth extra scrutiny.
It's a limitation, not a dealbreaker
Understanding hallucination isn't a reason to avoid AI tools — it's what lets you use them well. Treat AI output as a strong first draft or a well-read collaborator's guess, not a verified source, and you get the real benefit without the risk.
Related reading
See How Does AI Actually Work? for why prediction-based generation leads to this behavior, and the Hallucination entry in our AI Glossary for a quick definition.
Why it cannot simply be fixed
The obvious question is why the model does not just say when it does not know. The answer is structural rather than a missing feature.
The system produces likely continuations of text. "I do not know" is rarely the likely continuation in the writing it learned from — reference material, articles, and answers overwhelmingly contain answers. A confident fabrication is statistically more typical than an admission of ignorance.
There is also no internal distinction between recalling and inventing. Both are the same operation: producing probable text. Nothing in the process separates a fact seen a million times from a plausible-sounding invention, which is why confidence is not a signal you can read anything from.
Training can reduce the rate — models have improved measurably — but a system whose method is plausible continuation will sometimes continue plausibly and wrongly. That is the trade-off, not a defect awaiting a patch.
Where hallucinations cluster
They are not evenly distributed, and knowing where they concentrate is more useful than a general warning.
- Citations, references, and links. The format is highly predictable, so plausible fakes are easy to generate. Assume every reference is invented until checked.
- Specific numbers, dates, and statistics. Precision reads as authoritative and is often fabricated.
- Niche topics. Thin training coverage means more gaps filled by pattern.
- Recent events. Beyond the training cutoff, unless the tool searched.
- Anything about you or your organization. It has no information and will produce something anyway.
- Questions containing false premises. Ask why an event happened and it will usually explain, whether or not the event occurred.
Working around it
Ask for sources and check them. Not because it will not invent them, but because checking is fast and catches most of it.
Watch for the vagueness shift. Answers that drift from specific to general often mark the boundary of what it actually has.
Ask the same question in a new conversation. Consistent answers are weak evidence in favor; contradictory ones are strong evidence against.
Give it the source material. A model summarizing a document you provided hallucinates far less than one answering from memory. This is the principle behind RAG and tools like NotebookLM.
Verify anything that would matter if wrong. That is the whole rule, and it is not a large burden once it is a habit.
How to Fact-Check AI Output covers the checking process in detail.