Why You Should Never Ask an AI to Verify Its Own Citations

Large language models frequently generate fabricated citations that appear legitimate, complete with plausible author names, journal titles, and DOI numbers that resolve to nothing. Because LLMs produce text based on likely word sequences rather than stored publication databases, they cannot distinguish a real citation from an invented one. Asking the same model to confirm a citation's validity is unreliable, as self-evaluation bias causes it to reaffirm what it already generated. Experts recommend verifying citations through external metadata APIs such as Crossref or arXiv, which can confirm whether a DOI actually resolves to a real paper. Beyond existence checks, credibility and fidelity of sources must also be assessed independently, since a valid DOI does not guarantee a journal's legitimacy or that a paper supports the claim attributed to it.