Plausible formatting is not proof of existence
A fabricated reference can look polished: familiar surnames, a plausible journal, believable volume details and a DOI-shaped string. Language models can continue bibliographic patterns without retrieving a record. People can also merge details from separate search results. Treat any reference not exported from a source you checked as an unverified lead.
Warning signs include a title that never appears exactly, shifting initials, impossible issue dates, page ranges outside the volume or a DOI resolving elsewhere. None proves that AI was involved. The scholarly task is to establish bibliographic and substantive accuracy, not speculate about how the error arose.
Resolve the DOI technically and semantically
A DOI commonly begins with 10. and can be resolved through https://doi.org/. Remove trailing punctuation when copying it. Failure may indicate a typo, incomplete registration or invention. Successful resolution proves only that the identifier exists.
Compare the landing page with the claimed source. If title, creators, year or publication differ, the DOI is misassigned. Do not alter digits until something vaguely relevant appears. Search title and creators independently, then document any correction.
Require several identity fields to agree
Check Crossref metadata or the appropriate specialist registry, followed by the publisher or repository record. Normalised capitalisation and online-first dates can explain minor differences. Different creators, title or study cannot. Save the conflicting records before deciding.
Open the full text or an authorised version and locate the precise proposition. A genuine article may not contain the attributed result, or may say the opposite. An abstract and search snippet cannot establish detailed methods, table values or pagination. Check version, corrections and retraction status as well.
Do not quietly bend a wrong reference into shape
Where title and DOI identify different works, establish which item genuinely supports the claim. The title may be right and identifier wrong, or only a thematically similar paper may exist. Read any candidate before substituting it. Similar keywords are not a scholarly match.
Record the old entry, error, replacement and affected passages. Search the manuscript and reference manager for every short citation. Correct claims derived from the fabricated summary as well as the bibliography. If no evidence exists, remove or qualify the proposition rather than manufacture support.
Use AI for search ideas with a strict evidence boundary
A language model can suggest concepts or search terms, but it should not become the authority behind imported references. Run those terms in library catalogues and disciplinary databases. Do not ask for supposedly exact quotations or page numbers from documents the system has not been given.
Export records from a DOI registry, library catalogue or subject database into the reference manager, then review them. Automated imports can mishandle title case, editions and names. The guide to identifying scholarly sources supplies the broader quality check after identity is confirmed.
Audit the bibliography according to risk
Start with references originating in generated answers, inherited notes or unclear bulk imports. Then examine central evidence, direct quotations and unusually perfect matches. Label each record with provenance and verification date. A sample from the remaining list may expose a systematic import problem.
Reconcile citations with the bibliography: every used work should appear and every listed work should be used where the style requires. Verify page locators against the text. Similarity checking cannot reliably determine whether a DOI is fabricated; registry and full-text verification serve that purpose.
Handle an unresolvable record transparently
Try title variants, creator identifiers, library catalogues and subject databases, and ask a librarian when the claim is important. Record databases and date. “Not found” means not verified by the documented search, not logical proof that no copy exists anywhere.
If another work quotes the missing item, that creates a secondary-citation situation rather than direct access. Disclose it accordingly or find independent evidence. Never use the real metadata of the secondary work while retaining a claim supposedly taken from the absent original.