An AI hallucination is content generated without reliable support from facts or the supplied source. A language model might invent a court case, cite a paper that does not exist, or give the wrong date in a smooth answer. The term describes the output. It does not mean the model sees or hears something as a person might.
Fluent text can hide a weak claim
Language models learn to predict likely text. That skill produces readable sentences, but the prediction process does not contain a built-in fact checker. A model may combine familiar names, titles, and formats into a citation that looks genuine. Confidence in the wording does not show confidence in the fact.
Unclear prompts can increase the risk. A request may assume that an event happened when it did not. The model may follow the assumption instead of challenging it. Gaps in training data, conflicting sources, and unusual subjects also make unsupported output more likely.
Not every wrong answer has the same cause
A stale answer may reflect an old training cutoff rather than an invented memory. A calculator mistake can come from weak arithmetic. A retrieval system may supply the wrong document. People often group all these failures under hallucination, but the fix depends on the source.
It helps to examine the path. Did the search find a relevant source? Did the prompt tell the model to stay within that source? Did the model copy the number correctly? Did later software format it badly? Clear labels lead to better repairs.
Checks should match the stakes
A made-up detail in a private story draft may cause little trouble. A false drug dose or legal deadline can cause serious harm. High-stakes uses need qualified human review and trusted primary sources. A warning under the text cannot replace those controls.
Users can check quoted phrases, publication records, links, names, and dates. A source link helps only when it opens a real page that supports the claim. Several copied websites may repeat the same original error.
Ways to reduce unsupported output
Retrieval-augmented generation can give the model selected documents at answer time. The application can require citations from those documents and decline to answer when the evidence falls short. Structured tools can supply exact prices, inventory, or calculations.
Developers can lower creative sampling settings for factual tasks. They can ask for short answers with direct evidence. Evaluation sets reveal repeated weak spots before launch. None of these methods removes the problem in every case.
The safest design separates fluent generation from trusted facts. The model can explain or format information. A database, verified document, or expert can establish the underlying claim.