Ask an AI system for a fact and it may reply with the confidence of an expert. Sometimes it can even supply a detailed citation for a paper that does not exist. The unsettling part is not merely that the answer is wrong. It is that the wrong answer can look so complete.
These errors are often called hallucinations. NIST uses the term confabulation for confidently presented false or erroneous content. The label describes an output problem; it does not mean the system experiences anything like a human hallucination.
Learning patterns is not checking every claim
Large language models learn statistical relationships in training data and generate sequences of tokens. That process can support impressive language, reasoning and knowledge tasks. But generating a plausible continuation is not identical to consulting a verified database for every sentence.
A model can learn the structure of a scientific reference without reliably knowing a particular reference. It may combine a plausible author, journal and title into something that looks authentic. The grammatical packaging can be excellent while the underlying claim is unsupported.
This is why polished writing is a poor reliability test. Style and truth are connected in good human communication, but a text-generating system can reproduce the style without establishing the truth.
Several things can go wrong
Training information can be incomplete, outdated or contradictory. A prompt can contain a mistaken premise. The model may also blend details from different contexts or fail to maintain consistency across a long answer.
Questions asking for obscure specifics create particular difficulty when the relevant information is missing. A system encouraged to provide an answer may produce a guess instead of acknowledging that limitation. The apparent confidence in the prose does not necessarily represent a calibrated probability of being correct.
Not every error is a fabricated fact. A system might misread a real source, apply a valid rule incorrectly or answer a different question from the one asked. These failures need different checks even when users group them under the same label.
Why search helps without solving everything
Retrieval can supply relevant documents at answer time. Tools can perform calculations or consult a current database. Those additions can reduce reliance on information stored in model parameters.
But the system still has to select the right material and interpret it correctly. A real link may support only part of a claim, or the answer may overstate what the source says. Citations are useful evidence only when they actually connect to the proposition being made.
A model repeating its answer with greater confidence is not independent verification. A more meaningful check compares the claim with an appropriate primary source or a result obtained through a suitable tool.
What a trustworthy answer looks like
Good systems can express uncertainty, ask for missing context and distinguish observed evidence from inference. Evaluations and careful deployment practices also matter, particularly where errors have significant consequences.
For readers, the key habit is to separate fluency from support. A precise date, quotation or reference deserves a traceable basis, however natural the paragraph sounds. AI can help organise and explain information, but the smoothness of its language does not certify the facts inside it.