Verify an AI-suggested citation before it enters the manuscript

1. Keep the suggestion separate from verified evidence

Save the proposed title, author list, year, journal, identifier and the exact claim for which the reference was suggested. Record which tool supplied it and which details about that tool are actually available. If the interface does not expose a model version, say so. A screenshot of an answer can document what was suggested, but it cannot substitute for the underlying scholarly source.

ICMJE assigns humans responsibility for submitted material involving AI and rejects AI-generated material as the primary source for a claim. In this proposed workflow, the author therefore owns the decision to cite. The assistant can help organize the review queue, but a lead remains unverified until the relevant source and claim have been checked. Do not remove the uncertainty label merely because a second chatbot repeats the same citation.

Reference: ICMJE: AI use by authors

2. Check that the reference identifies the intended work

Follow a supplied identifier through the appropriate trusted service and compare the resulting title, authors, publication venue and dates with the suggested reference. Crossref exposes metadata deposited by members and trusted sources and provides DOI-record lookup. That is useful identity evidence for records it holds; it is not a finding that the study is correct or that every sentence attributed to it is supported.

If a title search finds a similar article, inspect it rather than replacing the identifier automatically. A conference abstract, protocol, preprint and full report may share authors and related titles while supporting different statements. Keep the exact version used. If the identifier belongs to another work, mark the proposed reference as a mismatch and search for the intended work independently. Do not repair it by inventing a DOI-shaped string.

Reference: Crossref: retrieve deposited metadata

3. Use different outcomes for different failures

The fictional team is drafting a paragraph about appointment scheduling. Lead A has a title about booking delays, but its supplied identifier points to an unrelated nutrition report. Lead B correctly identifies a report describing time to booking, while the AI sentence claims reduced hospital admissions. Lead C has a plausible title but no source has been retrieved. These are three different evidence states, not one generic “bad reference” category.

Keep identity and claim support in separate columns. A real paper can be cited for something it never studied. Conversely, inability to retrieve a source today does not prove that it does not exist. The table includes no real DOI or PMID, so it cannot be mistaken for a bibliography ready to paste into a manuscript.

On small screens, scroll the table horizontally, or focus it and use the arrow keys.

4. Read enough context to support the sentence

Read the methods, relevant result and limitations needed for the proposed statement. Record the page, section or other usable location. A keyword appearing in an abstract does not establish that the study measured that outcome. Check population, exposure or intervention, comparator, time point, direction and uncertainty. If only an abstract is available, limit the claim to what that abstract actually reports and record the access limitation.

In Lead B, the supported subject is a recorded booking interval. The supplied fictional facts contain no admissions outcome, effect estimate or causal analysis. A cautious replacement can say that the report describes time to booking, but it cannot announce a reduction in admissions or better health. A fluent paraphrase is still wrong if it silently changes the outcome or converts a description into a causal effect.

5. Document the search that actually happened

If a database search is needed, write its purpose before running it. A targeted search to locate one reference differs from a systematic search intended to identify all eligible studies. Record the database and interface, execution date, exact query, filters, returned records considered and the source selected. Keep an unsuccessful attempt too, with the actual error or limitation expressed without inventing a zero-result count.

PubMed’s help explains that Search Details shows query translation and warnings. For systematic reviews, PRISMA asks authors to report information sources, search dates and full strategies. In this guide, the database-execution fields remain blank because no PubMed results query was run. A team using the worksheet should fill them after execution; a suggested query in an AI answer is not evidence of execution.

Reference: NLM: PubMed search help · PRISMA 2020: expanded reporting checklist

6. Close the loop in the manuscript

Before submission, compare each in-text citation with the final reference list and the actual passage it supports. Ensure that deleting a mismatched lead did not leave its unsupported sentence behind. Check whether the source has a relevant correction or retraction notice, and record what was examined. If a source remains unresolved, make a deliberate editorial decision rather than allowing a polished reference manager entry to hide the gap.

Preserve a concise internal trail: suggested reference, verified identity or unresolved status, inspected passage, decision and reviewer. Keep licensed full texts and private annotations in authorized storage. Report AI assistance as required by the target journal using actual tool use. The completed output is a supported manuscript statement and traceable citation, not a large list of impressive-looking references.

Questions for a practical final check

• Exact AI suggestion and associated claim retained
• Identifier checked against the intended work and version
• Relevant text read with outcome and design intact
• Unretrieved source distinguished from a proven mismatch
• Query execution recorded only after it happened
• Final sentence and reference list checked together