Reference verification workflow
A plausible citation is still only a lead
A language model may propose synonyms, authors or article titles, but it is not a bibliographic authority. Even a real DOI can point to a different work. Maintain two distinct collections: a candidate log for unverified leads and a reference library containing only confirmed publications. That boundary prevents persuasive inventions from leaking into footnotes.
Begin with title fragments and author combinations. Then use your library catalogue, a relevant subject database, Crossref metadata and the publisher record. Search-engine results can help locate a work but do not alone establish its scholarly status. Where several manifestations exist, record preprint, accepted manuscript and version of record separately.
Verify metadata and claim as separate tasks
An accurate title does not prove that the publication supports the statement attributed to it. First establish identity: author names, date, journal or repository, volume, issue, pagination and DOI. Then open the source and read the relevant passage in context. Note population, design, period and limitations. A speculation in the discussion must not be rewritten as an observed result.
Existence is followed by source criticism
A confirmed publication may still be unsuitable. Assess publication type, date, method, conflicts, corrections or retraction, and relevance to your setting. One highly cited article does not replace a balanced search. Record databases, terms, filters and reasons for inclusion so that contrary evidence is not silently excluded.
A systematic search cannot be replaced by a conversation. Search strings must be tested in the databases actually used and logged with date, hit count and selection process. The systematic search protocol provides a dedicated structure.
Describe the AI role proportionately
A concise statement might read: “On 16 August 2026 I used [tool] to propose English synonyms and possible starting publications. I independently searched every reference in [databases], compared metadata with authoritative records and checked relevant claims in the full text. Unconfirmed suggestions were excluded.” Add the affected chapter and scale.
The AI transparency hub connects search records with model and prompt evidence. Why AI output is not a source explains the conceptual boundary. The academic writing overview places discovery within the full research process.
Pre-citation quality check
- Is the record uniquely identified by DOI, catalogue entry or stable URL?
- Do all metadata agree with the original?
- Have you read the passage that supports your proposition?
- Did you check for corrections or retraction?
- Is the cited version explicit?
- Does the method justify the use you make of it?
- Were relevant counterpositions sought?
- Is every unresolved candidate excluded from the bibliography?
Keep a short note alongside the library record giving the verified location and passage. It makes page checking easier and prevents memory from turning confirmed existence into imagined support.
Deduplicate versions, not just identifiers
The same research may appear as preprint, conference paper and later journal article. Compare author group, title, abstract and publication history rather than DOI alone. Select the version relevant to your claim and never count manifestations as independent evidence. Keep a short exclusion note such as “similar title, different population” or “DOI belongs to another work”. It prevents repeated checking and accidental re-import of a false candidate.
Sources
Give unresolved candidates a firm boundary: if their record and original cannot be confirmed before the search closes, exclude them from manuscript and bibliography. Record unsuccessful search routes so that the same invented item does not return under a slightly altered title. For important inaccessible work, use library services or a legitimate alternative version. Only the version actually read determines the claims you can responsibly make.