A hallucination is when an AI model states something false with complete confidence. The word makes it sound like a malfunction. It is not — it is the system working exactly as designed, and understanding why is the only reliable way to protect yourself from it.
Why it happens
A language model predicts what text should come next, over and over. That is the entire mechanism.
It has no database to check against, no internal flag for "I don't know this", and no way to distinguish between recalling something and constructing something plausible. Both feel identical from the inside, because internally they are identical.
So when it lacks the information, it does not stop. It produces the most plausible continuation — which is a confident, well-formed, wrong answer.
The model is not lying. Lying requires knowing the truth. It is doing what it was built to do with insufficient information.
Why confidence tells you nothing
This is the part that catches people.
The fluency and certainty of an answer are unrelated to its accuracy. A model generates confident-sounding prose because confident-sounding prose is what its training data looks like — not because it has verified anything.
Asking "are you sure?" does not help much either. The response is itself generated text, not a report from an internal confidence meter.
Where it happens most
Ranked by how often it bites people:
- Citations and sources. The highest-risk output there is. A plausible reference is precisely what the mechanism produces best, and fabricated citations have ended up in court filings and published papers.
- Specific numbers. Statistics, dates, prices, measurements — especially with a confident source attached.
- Recent events, without web access. Training data has a cutoff and the model has no awareness of it.
- Niche or specialist topics where training data was thin.
- Details about your own organisation. It knows nothing about your refund policy and will invent a reasonable one.
- Library and API methods in code. Confidently calling functions that do not exist.
What actually reduces it
Ground it in material you supply. The single most effective technique. Paste the document and add: "Answer only from the text below. If the answer isn't in it, say 'not covered'." This works because you have removed the gap the model would otherwise fill.
Use a model with live search for anything current. Search-connected answers are constrained by retrieved sources, and you can check the links.
Ask for the source separately, then verify it exists. Do not accept a citation as evidence of itself.
Cross-check in a second model. Two models inventing the same specific false detail is much less likely than one doing it.
Lower the stakes of being wrong. Use it for drafting and analysis, where you will read the output carefully anyway, rather than for facts you will paste onward unchecked.
What does not work
Better prompting alone. "Be accurate", "do not hallucinate", "only tell me true things" have little effect. The model is not choosing to be inaccurate.
Bigger models. Larger and newer models hallucinate less, not never. The rate falls; the failure mode does not disappear.
Trusting hedged language. Models hedge stylistically, not because they detected uncertainty.
Asking it to check its own work. Sometimes catches an error, sometimes confidently confirms the wrong answer.
A practical rule
Ask: if this were wrong, when would I find out?
If you would notice immediately — a draft you are about to edit, code you are about to run, a summary of a document you have open — the risk is low, and this is where AI is most valuable.
If you would find out much later, or never — a statistic in a report, a citation in a document, a legal or medical claim — verify every single one, or do not use it there.
More on the mechanism: what GPT is and how it works.
Common questions
What is an AI hallucination? When a model states something false with complete confidence. It happens because the model generates plausible text rather than retrieving verified facts.
Why does AI make up sources? A plausible-looking citation is exactly what a next-word predictor produces well. It has no mechanism for checking whether a source exists.
Can hallucinations be fixed? Reduced substantially, not eliminated. Grounding in supplied documents and live search are the effective mitigations.
Do newer models hallucinate less? Yes, meaningfully. They still hallucinate.
Does telling it "don't hallucinate" work? Barely. The model is not choosing to be wrong, so instructing it not to be has little effect.
How do I know if an answer is hallucinated? You cannot tell from the answer itself — confidence and fluency are no guide. Verify anything that matters, particularly numbers and citations.
Which model hallucinates least? Models with live web access are more likely to be current and are constrained by retrieved sources. That is different from being accurate, and no model is reliable enough to skip verification.