An AI hallucination is an answer that a language model presents as fact but that is not true — an invented citation, a misremembered date, a function that does not exist in the library, a court case nobody filed. The word is a little generous to the machine: nothing is being perceived. The model is doing exactly what it was built to do, and a wrong answer is produced by the same process as a right one. The term is also used for a second, narrower failure: an answer that contradicts a document the model was given, even where the statement happens to be true of the world.
Why models hallucinate
A large language model is not a database with a search box in front of it. It is a system that predicts plausible continuations of text, one piece at a time, from patterns learned during training. When the pattern and the fact agree, the output is correct. When the training data was thin, contradictory, or simply absent for your particular question, the model still produces the most plausible-looking continuation — because producing one is the whole job. There is no step where it checks a claim against a source, and while research shows models do carry internal signals that correlate with being wrong, none of them reaches you as a reliable "I do not know this."
That is why they are most common in a few predictable places: precise details a model has no reason to have memorised (page numbers, case citations, API parameters, dates), questions about events after its training cutoff, and anything where the shape of a correct answer is well known but the content is not. A fabricated citation looks exactly like a real one because the format is what the model learned best.
Confidence is a poor signal
The most practical thing to know is that tone is close to useless as a check. Models are trained to write like a competent, helpful expert, and they write that way whether or not the underlying claim is sound. A hedge ("I believe", "as of my last update") largely reflects patterns in the training text rather than a measurement of the model's own reliability. Research can extract uncertainty signals that beat chance at predicting when an answer is wrong, but those are not what the prose is showing you — so reading a confident paragraph as evidence is one of the easiest ways to let a hallucination past.
How to catch one
Three things work, in rough order of effort. Check the primary source — if a claim has a citation, open it; fabricated references collapse immediately. Ground the question by giving the model the document or using a mode that searches the web, so it is summarising something real rather than recalling. And ask more than one model: independently trained systems fail in different places, so a claim that two or three of them state the same way has cleared a bar that one model's confidence never establishes.
That last one is what the AI Council automates. One question goes to several models from different providers at once, their answers are shown side by side, and a referee marks the specific claims where they disagree — so instead of one fluent paragraph you get the map of which parts are contested. Disagreement is the useful output: it points at exactly the sentence worth checking yourself.
The honest limit: agreement is not proof
Cross-checking narrows the problem; it does not close it. Models are trained on overlapping data, so they share some of the same gaps and some of the same mistakes. Ask several of them about a widely repeated myth and they will often agree — on the correct debunking, which is fine, but they can just as easily agree on a popular error, and the agreement will look identical either way. Consensus raises confidence; it never establishes truth. For anything load-bearing — medical, legal, financial, or something you are about to publish — a source still has to be checked by a person. What several models are genuinely good at is telling you where to spend that effort.
Frequently Asked Questions
- Can hallucinations be fixed by a better model?
They can be reduced, and later flagship models have measured lower hallucination rates, though it is not a simple rule that newer or larger is always more truthful. The mechanism that produces them is the same mechanism that produces useful answers — predicting plausible text. Grounding a model in real sources tends to help more than reaching for a bigger model, and no current system removes the problem entirely.
- Does web search stop hallucinations?
It helps a lot and does not eliminate them. A grounded model is summarising retrieved pages rather than recalling from memory, which removes the most common failure — but it can still misread a source, over-generalise from one, or cite a page that does not support the sentence it is attached to. Opening the link remains the check.
- Why do two AI models give different answers to the same question?
Different training data, different fine-tuning, different cutoffs, and sampling that is not deterministic. On settled facts they usually converge, though not always — an ordinary generation error can split them on something uncontroversial. Divergence is a flag worth following rather than a diagnosis: it often means the question is contested, poorly specified, or outside what either model reliably knows.
- Is a hallucination the same as a lie?
No. Lying requires knowing the truth and choosing otherwise. Lying needs an intent to deceive, and the model is not weighing a truth it holds against the answer it gives — it produced the most plausible text it could. The practical consequence is the same for the reader, which is why the checking habit matters more than the terminology.
Related terms: Context Window · System Prompt · all glossary terms