Why do AI chatbots invent citations?

Every researcher who has asked a chatbot for sources has met the invented reference. It looks right, it is formatted perfectly, and it does not exist.

Because a language model writes text shaped like a citation; it does not look one up. Trained to produce plausible text, it can generate a convincing author, title and year whether or not the source exists. Retrieval fixes half of this by giving the model real passages to read. The other half needs a gate: every claim checked against the passage it cites, and a refusal when the sources contain no answer.

Why models invent sources.

A language model has no index of the world to consult, a property behind every hallucination. It has a statistical memory of what text tends to follow other text, and "the study, published in a journal, found" is a pattern it has seen many times. So it continues the pattern. The result is not a lie in any useful sense; there is no intent behind it. It is a sentence shaped exactly like a citation, generated with the same ease as any other sentence, and no amount of careful prompting changes what the pattern continues into. That is also why the invented ones feel so credible: the fluency is real, only the referent is missing.

Both citations are built the same way, down to the one field that can be checked: the locator, in red, is the part a reader can follow. The first is a record that exists and its line ends at a document. The second is not a real record, drawn here as an example, and its line has nothing to arrive at, so it leaves the frame.

Why retrieval alone does not fix it.

Retrieval-augmented generation changes the setup: instead of answering from memory, the model is handed actual passages from your documents and asked to answer from those. The invented-source problem mostly disappears, because the model now has real material to cite. What does not disappear is the accuracy problem. A model can still quote a passage slightly beyond what it says, merge two authors into one claim, or cite the right document for the wrong sentence. The answer is grounded in real sources and still partly wrong.

What actually fixes it.

The fix is structural, not a matter of asking the model nicely. After the model writes its answer, a separate verification pass compares each claim against the exact passage the answer attaches to it. Claims the passage does not support get removed or rewritten; a question the sources cannot answer gets a refusal instead of a guess. The full stack, retrieval plus citations plus verification, is laid out on how to stop an AI making things up. That gate is the difference between an answer that carries citations and an answer that earns them, and it also settles how such answers get cited downstream: the sources get cited, the tool gets credited. You can see it in any grounded tool by doing one thing: click the citation. A real one opens the passage, and the passage says what the answer said it would. That test works on cloud assistants and local ones alike, and it is the first thing to run on any tool before it gets to touch your research.