The Mechanism, Precisely
A large language model generates text, including citations, by predicting the next most statistically likely token based on patterns learned during training — it does not query a live database of academic papers, books, or articles when asked to provide a citation, unless specifically connected to a retrieval tool. When asked for a source on a given topic, the model produces a reference that matches the structural pattern of citations it saw during training: a plausible author name, a real or real-sounding journal, a year that fits the topic's timeline, and often a DOI-formatted string.
The model has no internal mechanism to flag this output as different from a citation to a source that actually exists in its training data, because both are generated by the identical underlying process — predicting plausible next tokens. This is the core reason hallucinated citations are so difficult to catch by reading: there is no qualitative difference in how confidently or fluently the model presents a real versus a fabricated citation.
Hallucination vs. Ordinary Citation Error
A human citation error — a wrong page number, a misspelled author name, an incorrect year — typically originates from a real source that was looked at, with an error introduced in transcription. Citation hallucination is categorically different: there is no real source being mistranscribed. The reference is generated whole, with every field constructed to be plausible rather than copied from anything that exists.
This distinction matters for how each is caught. A transcription error is often catchable by checking the cited source directly — the source exists, just with one field wrong. A hallucinated citation has no source to check against at all, which is why structural and identifier-based detection methods are necessary rather than simple proofreading against the original.