An AI hallucination is a fluent, confident answer that the model generates without any grounding in fact or source — it is not lying, and it is not retrieving a wrong fact, it is producing text that reads as authoritative while having no factual basis at all.
That is the single most important thing to understand: the model is not checking whether what it says is true, it is predicting what words would plausibly come next. The confidence in the tone is a property of how the sentence is built, not evidence that the sentence is correct.
To see why this happens, you have to look at what these systems actually do. A large language model is trained on enormous amounts of text to predict the next token — a token being a chunk of a word — given everything that came before it. During training it is rewarded for producing sequences that look like the text it has seen.
Nothing in that objective says "verify this against a source" or "say 'I don't know' when unsure." So when you ask for a legal citation, a Python method name, or a statistic, the model produces the shape of an answer: a plausible-looking citation, a method name that fits the naming conventions of the library, a number in a realistic range.
Fluent, well-formatted, and potentially invented. This is why hallucinations are so hard to spot by eye — the wrong answer often looks cleaner and more confident than a correct one, because confidence is exactly what the training rewarded. A model that hedged every sentence would score poorly; a model that asserts smoothly scores well, whether or not the assertion is true.
There is a second layer too: models are trained to be helpful and to complete the task you gave them. If you ask for five sources, refusing to produce them feels like failing the task, so the model tends to produce five sources rather than say it does not know. That helpfulness pressure makes fabrication more likely precisely when you most want a refusal.
According to our learn page on AI hallucinations, this is a structural behavior of how the tools are built, not a bug that a future version will simply patch out.
A concrete example makes this vivid. Suppose you ask a chatbot: "Give me a real court case holding that a company can be liable for an AI chatbot's false statements about its own product." A hallucinating model may reply with something like "See *Reyes v.
Meridian Analytics* (9th Cir. 2023), 47 F.4th 1122," complete with a plausible court, circuit, year, volume, and page range. Every element is fabricated.
The citation format is real, the reporter abbreviation is real, the page range looks right — and the case does not exist. This is not a hypothetical pattern; fabricated citations are one of the most commonly reported hallucination types because legal writing has such rigid, checkable formatting that a model can imitate the form perfectly while inventing the content.
The same thing happens with API documentation (a method like `client.fetch_embeddings_batch()` that sounds exactly like the real library's naming style but was never implemented) and with statistics (a percentage that sits in a believable range for the topic but traces to no study). The practical check is simple and you should run it every time the answer matters: take the single most load-bearing fact — the case name, the method signature, the number — and try to verify it against a primary source outside the chat.
Look the case up in a legal database, run the method in a scratch file, search the statistic on the original publisher's site. If you cannot find it independently, treat it as unverified no matter how confident the phrasing was. A useful rule of thumb: the more specific and checkable a claim is (a number, a date, a name, a citation), the more it is worth verifying, because specificity is exactly what the model is best at faking.
What actually reduces hallucinations, and what does not, is worth being honest about. Retrieval-augmented generation — where the tool first searches a document set or the web and then answers using those passages — genuinely helps, because the model is now copying from supplied text rather than free-associating.
But it reduces hallucinations; it does not eliminate them. The model can still misread the retrieved passage, blend two sources, or answer from memory when retrieval returns nothing useful. Lowering the temperature setting makes output more repetitive and slightly less prone to wild invention, but it does not make the model truthful.
Asking the model to "only answer from the provided documents" helps at the margins and fails when the documents do not contain the answer, because the helpfulness pressure pushes it to answer anyway. The honest limits are these: there is no setting that turns a next-token predictor into a fact-checker; every mitigation is a risk reduction, not a guarantee; and the failure mode is worst exactly where you most need reliability — niche topics, recent events, precise numbers, and anything with a formal citation format.
The practical posture is to use these tools for drafting, structuring, and explaining, and to treat every specific factual claim as a draft that needs a source before it goes anywhere that matters. If you want to go deeper on telling when output is wrong, our guide on why AI makes things up and how to spot it walks through the warning signs in more detail.