Journals can screen references at submission with a desk check of each manuscript’s reference list before peer review. The desk check covers four problems: citations to retracted papers, references that match no real publication, incomplete or broken references, and claims their sources do not support. Proofig AI’s RefGuard flags all four, and editors decide what goes back to authors.
Why screen references before peer review?
Peer reviewers judge the science; they are rarely asked to resolve each entry in a bibliography. Reference problems therefore often surface after publication, when fixing them takes a correction or a retraction.
A correspondence in The Lancet audited 97.1 million references in open-access biomedical papers and estimated that the rate of fabricated references rose more than twelvefold, from about 4 per 10,000 papers in 2023 to 56.9 per 10,000 in early 2026. The authors note that reference verification is not standard in peer review and recommend that publishers build automated verification into submission workflows before review begins. In their words, “the barrier to adoption is institutional rather than technological.”
What should the desk screen in a reference list?
Four checks cover the problems that matter most to an editor:
- Retracted citations. A cited paper that has been retracted, used as if its findings still stand. Citing authors are rarely alerted when a retraction lands (see why citing retracted papers happens).
- Unverifiable or hallucinated references. Entries that match no real publication. AI writing tools invent references with real author names, real journals and plausible years that pass a visual check (see how hallucinated references arise).
- Broken DOIs and metadata. DOIs that do not resolve or that point to a different paper, and authors, years or journals that disagree with the record. Crossref member journals must link each item in a reference list with its DOI, so broken identifiers also cost time in production.
- Claims without support. Statements that need a source but have none, and citations whose source does not back the claim or contradicts it.
How should editors triage reference flags?
COPE’s 2026 topic discussion, When citation integrity is questioned, suggests that editors weigh four things: the nature of the problem (partly genuine or wholly fabricated), its location (a meta-analysis or a passing remark in the discussion), its category (a reference that cannot be found, or a real one used inappropriately) and its scale (one or two references, or a large share of the list).
A three-tier triage follows from that:
- Hold before review. A retracted paper cited as evidence for a result, a method or the study’s rationale. References that match no publication, especially several of them, or any inside methods sections, meta-analyses and reviews. In the Lancet audit, review articles had a fabrication rate 57% higher than other paper types.
- Query the authors. Citations that do not fit their claim, claims with no source, and retracted papers whose role in the argument is unclear.
- Fix in revision. Incomplete or wrong metadata, DOIs with typos, dated sources where newer evidence exists, and paratext, such as news items or web pages, cited as research.
What should go back to the authors?
A precise request gets a better revision than a general note. For each flagged entry, send:
- the reference number, the problem and the evidence, such as the retraction date and reason or the article a DOI really resolves to;
- the action needed: remove or replace a retracted source, or keep it with a clear statement of the retraction and the reason it is still cited;
- for references that cannot be verified, a request for a working DOI or a copy of the source (COPE’s discussion suggests this when DOIs do not resolve outside the authors’ home country);
- a short account of how the reference list was compiled, including any AI tools used to find or format references;
- for unsupported claims, a better source, a corrected citation or a more modest claim.
What do COPE and ICMJE expect from editors and authors?
Responsibility for references sits with authors. The ICMJE Recommendations make authors responsible for citing accurately and for checking that no reference cites a retracted article, except when referring to the retraction. ICMJE’s section on AI use by authors adds that humans must ensure appropriate attribution of all quoted material, including full citations.
Editors own the process. ICMJE asks editors to query authors about whether and how AI was used, and warns against uploading submitted manuscripts into AI systems where confidentiality cannot be assured without the authors’ explicit permission.
COPE’s retraction guidelines ask for prompt notices to limit harm, including further citation of erroneous work. A desk screen closes that loop from the submission side.
How does reference screening fit with image and full-paper checks?
References are one layer of a submission check. Figures need their own screen for duplication, manipulation, image plagiarism and AI-generated images, which is Proofig AI’s image integrity work. Statistics, methods, ethics disclosures and authorship details need a reading of the whole paper, which Octym provides.
Running these checks at one intake point gives the handling editor one picture of the manuscript and helps reveal patterns. In the Lancet audit, some fabricated references clustered in ways consistent with paper-mill activity, including the same pair of authors on 11 papers in a single journal. A fabricated reference list next to a reused figure tells an editor far more than either flag alone.
Where does RefGuard fit at the editorial desk?
RefGuard is Proofig AI’s reference integrity product, and it flags these problems in a manuscript’s reference list before peer review. It accepts text-based PDF and Word files of up to 150 pages and compares each reference with the scholarly publishing record to confirm the article is real.
- Hallucinated references. References that match no real publication are flagged, so the desk sees them before peer review.
- Retracted citations. RefGuard flags citations to retracted papers with the reason, the date and the type of notice, drawn from a proprietary retraction database.
- Claims and citations. RefGuard flags statements that need a source but lack one, and citations whose source does not fit the claim or contradicts it.
- Incomplete or broken references. RefGuard points out incomplete or broken references, such as a wrong year or a missing author, so they stand out.
- The shape of the list. RefGuard gives Quality, Impact, Recency and Diversity scores, shows publication years, citation counts and open-access share, and notes references 15 or more years old and paratext. The list can be sorted, filtered and exported as CSV.
RefGuard is available on its own, inside PubShield, Proofig AI’s integration hub that returns integrity checks to one dashboard, and as the reference checks in Octym, which reviews the entire paper with 26 checks. Proofig AI for publishers shows how image and reference screening fit an editorial office. RefGuard’s flags are signals for an editor to review, not verdicts, and authors stay responsible for their references.
Frequently asked questions
Should a journal reject a manuscript that cites a retracted paper?
Usually not: a citation to a retracted paper is rarely grounds for rejection by itself. ICMJE accepts such a citation when the manuscript refers to the retraction itself. The problem is a retracted study used as evidence for a result, a method or the rationale. Ask the authors to remove or replace it, or to state the retraction and justify keeping it, then judge whether the conclusions still stand. RefGuard shows the retraction reason, date and type of notice, so the editor has the facts to decide.
When should unverifiable references go to a research integrity team?
Unverifiable references should go to a research integrity team when they form a pattern. One reference that cannot be found is usually a query for the authors, who can supply a working DOI or a copy of the source. Several unverifiable references in one list, or any in a meta-analysis, call for escalation. So do problem citations that recur across articles, which COPE’s citation integrity discussion says may signal systematic manipulation. RefGuard’s flags give an integrity team a documented starting point.
What should a journal’s author guidelines say about reference checks?
A journal’s author guidelines should state that references are screened at submission and explain what happens to flagged entries. Useful points include DOIs given as full links, disclosure of any AI tools used to find or format references, verification of each source against a bibliographic database or the original, and a rule on retracted papers that matches ICMJE’s. COPE’s discussion advises journals to have policies on declaring AI use, including for references. Proofig AI’s RefGuard then applies the screen consistently.
