Screening manuscript references for fabrication and retraction

Fabricated references now reach journals, not just coursework. Topaz, Zhang, et al. (2026) report that their audit of 2,471,758 PubMed Central articles found 4,046 references to works that do not exist, spread across 2,810 peer-reviewed papers, with the rate rising more than twelvefold from 2023 to early 2026. Retracted papers keep their places in reference lists too, and the editorial office is often the last point at which either can be stopped before it enters the record.

This guide, published by EdCitation, is for editors and editorial-office staff. Every rule in it was read at its source on 2 October 2026: the ICMJE Recommendations (updated January 2026), COPE's guidance, and the policies of Elsevier, Springer Nature, PLOS and Taylor & Francis. It ends with a screening procedure and a test in EdCitation's free Verify references, which looks each entry up in the publisher's record, flags retractions, and never writes a reference for anyone.

The reviewer's half of this work is in checking the references in a manuscript as a peer reviewer; coercive citation has its own guide, and so do retraction notices.

What should an editorial office screen references for?

Seven things. A lookup finds the first four in minutes; the last three need someone who reads the manuscript.

What do the ICMJE and COPE ask of editors?

Both put the duty of accuracy on authors and the duty of vigilance on editors.

The ICMJE Recommendations, January 2026

Authors are responsible for citing accurately and should be able to attest that each reference supports its statement. References should be verified against a bibliographic source such as PubMed or the originals, and none may cite a retracted article except to discuss the retraction; for MEDLINE journals, PubMed is the authoritative source on retractions (International Committee of Medical Journal Editors [ICMJE], 2026).

On AI, citing AI-generated material as the primary source "is not acceptable", journals should require disclosure at submission, and editors should stay alert to authors who do not disclose AI use when asked. Editors and reviewers should not upload a submitted manuscript into an AI system where confidentiality cannot be assured, without the authors' explicit permission (ICMJE, 2026).

What COPE says about questionable references

COPE's June 2026 Forum discussion, "When citation integrity is questioned", introduced by a Council member and a COPE adviser, advises weighing four things before acting (Committee on Publication Ethics [COPE], 2026a):

• Nature: is the reference partly genuine or wholly fabricated?
• Location: does it sit in a meta-analysis or a discussion paragraph?
• Category: is it unfindable, or real but used for something it does not say?
• Scale: is it one or two entries, or a large share of the list?

It adds that DOIs which resolve to another article can indicate AI was used to find references, that authors who blame a formatting tool can be asked for a time-stamped draft, that wholly fake references may mean the findings are not grounded in evidence, and that problem citations recurring across articles belong with the publisher's integrity team. COPE's March 2026 discussion on suspected AI lists hallucinated references, references that do not support their statements, and formulaic patterns such as many references from one year among its warning signs (COPE, 2026b).

After publication, COPE's retraction guidelines, revised in August 2025, apply: an editor may retract when they no longer have confidence in the results and conclusions (COPE Council, 2025). In a 2026 case, COPE Council quoted them as listing undisclosed AI involvement among the possible reasons, and warned editors not to let an AI-detection tool judge for them (COPE Council, 2026).

What do publishers' policies say about retracted and AI-generated references?

Several large publishers now screen references at submission, and their policies say what must happen when a cited paper has been withdrawn.

Citing retracted work

PLOS tells authors that submitted research "should not rely upon retracted research". A retracted paper may be cited for historical context only, with its status stated in the text and the reference list, and a work retracted or corrected before publication, if still cited, is cited together with its notice (PLOS, n.d.-b).

Elsevier's reference check in Editorial Manager marks entries "retracted" or "not validated". Editors are told to look for a reasonable explanation in the manuscript or cover letter and, failing one, to ask the authors to confirm their research is not compromised and explain the reference in the text; citing the retraction notice is called best practice (Elsevier, 2025). "Not validated" can simply mean the source is not indexed in Scopus, PubMed or Crossref: no record is not the same as no paper.

AI-generated content and references

Springer Nature (2025) launched an in-house tool in April 2025 that judges the relevance of every reference in submissions to almost all its journals and books; when several look irrelevant, its Research Integrity Group checks the manuscript by hand and decides whether to withdraw it.

Taylor & Francis (n.d.) lists lack of attribution of "ideas, quotes, or citations" among the risks of generative AI, requires authors to name the tool, version and use, and says editors must not upload unpublished manuscripts into generative AI tools.

PLOS makes authors responsible for citing all relevant sources and no irrelevant ones, and forbids editors and reviewers to upload submission content to generative AI tools, web platforms, databases or other services, while it may use web-based tools itself during submission checks (PLOS, n.d.-a). At such a journal, the reference screen belongs to the editorial office's own systems, not to any outside checker, EdCitation's included.

How often do fabricated and retracted references reach journals?

Rarely per paper, more often every year, and peer review removes only some of them.

• Preprints against journals. Topaz, Zhang, et al. (2026) scanned 42,249 medRxiv preprints posted from January 2023 to March 2026. Two researchers, blind to each other's verdicts, reviewed 184 flagged candidates and confirmed 157 fabricated references in 104 preprints: 37.2 per 10,000 preprints, against 16.4 per 10,000 in the team's PubMed Central audit (Topaz, Roguin, et al., 2026). Twelve of the affected preprints were later published in journals with the fabrications intact. The authors note that they cannot estimate how many fabrications they missed, and recommend automated reference checks at submission.
• Across four repositories. Zhao et al. (2026), in a preprint, audited 111 million references in 2.5 million papers on arXiv, bioRxiv, SSRN and PubMed Central, and put the number of hallucinated citations in 2025 alone at a conservative 146,932. They conclude that preprint moderation and journal processes catch only a fraction.
• Retracted paper-mill articles. Tang and Cai (2025) checked 200,000 systematic reviews published from 2013 to 2024. Of these, 299 (0.15%) built at least one retracted paper-mill article into their evidence synthesis, and 124 of the 385 citations (32.2%) were made after the retraction.

How often chatbots invent references, model by model, is in what studies say about AI-fabricated references, and the long record of retracted papers cited as sound is in retracted papers still being cited.

What we could not confirm: we could not open the full text of the Lancet audit, so its figures here are those its authors give in their later letter. COPE's 2025 retraction guidelines would not download, so the wording on AI comes from COPE Council's quotation of them in the 2026 case.

How does an editorial office screen references, step by step?

Screen once at submission and again before proofs, cheapest checks first, so that reading time goes where the lookups point.

1. Write the policy down. COPE (2026a) says journals need policies on AI use, including for finding and formatting references, and on verifying sources editors cannot reach.
2. Decide which tools may see a submission. Follow the publisher: PLOS and Taylor & Francis restrict uploads, and the ICMJE asks for confidentiality.
3. Look every reference up at submission, before review, with the publisher's own checker or, where policy allows, Verify references. The ICMJE makes verification the authors' job, so the instructions to authors can ask them to run a check first.
4. Sort the results: retracted, no record, DOI leading elsewhere, detail error, could not check. A check that could not run is evidence of nothing.
5. Search each no-record entry by hand in Crossref, PubMed, OpenAlex and Google Scholar. Theses, reports and older or non-English works often escape the indexes, as my reference cannot be found explains.
6. Weigh nature, location, category and scale, as COPE (2026a) suggests.
7. Write to the authors neutrally, quoting the journal's own guidelines, as COPE Council (2026) advises. Ask for a DOI, a link or a copy, and how the list was compiled.
8. Screen again before proofs. Questionable references added after review can signal a reviewer's coercive citation (COPE, 2026a).
9. Look across submissions. The same suspect references in several manuscripts belong with the publisher's integrity team.
10. After publication, choose the notice that fits the effect on the findings: correction, expression of concern or retraction (COPE Council, 2025).

Wording that accuses no one

State the fact, the place and the request, and leave out any guess about why:

• "References 14 and 27 could not be found in Crossref, PubMed or OpenAlex. Please send a DOI, a link or a copy of each, or replace them."
• "The DOI given for reference 31 leads to a different article. Please correct the entry and tell us how the reference list was compiled, including any AI tools used."

What Verify references returned on a five-entry test list

On 2 October 2026 we pasted five references into Verify references: the retracted 2020 Lancet hydroxychloroquine paper by Mehra and colleagues, with its DOI; Tang and Cai's study, with its DOI; Topaz and colleagues' letter, without one; Bik and colleagues' image-duplication study, dated 2017 instead of 2016; and an entry we invented, carrying the real DOI of a 2025 Learned Publishing article. Our invented entry, exactly as pasted:

Okafor, L. N., & Brandt, H. E. (2024). Screening submitted manuscripts for non-existent references: A survey of 412 journal editorial offices. Learned Publishing, 37(2), 118-131. https://doi.org/10.1002/leap.1667

The verdicts, quoted as returned:

• Mehra and colleagues (retracted): verified and flagged, "The DOI resolves to this record and the title matches. This paper has since been retracted."
• Tang and Cai: verified, "The DOI resolves to this record and the title matches."
• Topaz and colleagues, no DOI: verified, "Title, year and first author match a published record."
• Bik and colleagues, wrong year: "check this", "The work was published in 2016, not 2017; the record has no date in 2017."
• Our invented entry: "check this", "The DOI is real, but it belongs to a different article from the one described."

The tally: 2 verified, 2 to check, 0 not found, 1 retracted, 0 unchecked. Pasted again without its DOI, the invented entry came back not found: "No publisher's or registry's record matches this reference." Given the retracted paper's DOI, Cite a source built the entry from Crossref's record, marked it retracted and listed the 2020 retraction notice.

What the run shows an editor

The borrowed DOI is the pattern COPE (2026a) describes, and a lookup that only asked whether the DOI exists would have passed it. Verify references compared the title behind the DOI with the entry and reported the mismatch; without the DOI, the same fabrication came back not found. Either way it goes to step 5 for a person to judge: the tool narrowed five entries to three that need a look, each with a reason.

Which EdCitation tools fit an editorial office?

For confirming that every entry is a real, published work that has not been retracted, EdCitation's Verify references is the best tool we know of: it looks each reference up in publishers' and registries' records and never writes one, where a chatbot asked to check a list can invent the answer. Above, it caught a retraction, a wrong year and a borrowed DOI, each with its reason.

• Verify references, free with no account: paste a list or upload a paper; each entry comes back verified, doubtful ("check this") or not found, retracted papers are flagged, and "could not check" is never shown as "not found". With the whole paper uploaded, in-text citations and entries are matched both ways.
• References from a file, part of Pro at $8 a month: upload a Word document or PDF; every reference is looked up, every in-text citation matched, corrections linked, and published sources recommended for a reference that cannot stand. Three checks a month come with a free account (pricing).
• Cite a source, free: builds an entry from a DOI, PubMed ID, title, ISBN or web address in APA 7, MLA 9, Chicago 18 author-date, Harvard, IEEE or Vancouver, screened for retraction.

EdCitation finds, cites and verifies, and writes no part of anyone's work, whether manuscript, review or decision letter: the judgement stays with the editor, as COPE and the ICMJE expect. Use it where your publisher's confidentiality rules allow, and point authors to it before they submit.

Quick questions

Should a journal check every reference in every submission?

No rule we read requires it, but the ICMJE makes authors responsible for verified, unretracted references, and Elsevier and Springer Nature already screen at submission. A lookup covers every entry; reading goes to the flagged ones.

Is a reference that cannot be found proof of fabrication?

No. Elsevier's checker marks a reference "not validated" when its source is not indexed in Scopus, PubMed or Crossref. Ask the authors for a DOI, a link or a copy before concluding anything.

Can authors cite a retracted paper?

Only to discuss the retraction or give historical context, with its status stated. The ICMJE and PLOS both say so. EdCitation's Cite a source shows a retraction when it builds the entry.

Can an editor paste a submission's references into an online checker?

Only where the publisher allows it. PLOS forbids uploading submission content to outside websites and databases. Otherwise use the journal's own system, or ask authors to run Verify references before they submit.

Should a paper with fabricated references be retracted?

It depends on what the references carried. COPE reserves retraction for work an editor can no longer stand behind: a fake reference in a discussion paragraph may need a correction, while fake studies in a meta-analysis can undermine its findings.

References

• Committee on Publication Ethics. (2026a). When citation integrity is questioned [Topic discussion]. COPE. https://publicationethics.org/topic-discussions/when-citation-integrity-questioned
• Committee on Publication Ethics. (2026b). When editors suspect AI [Topic discussion]. COPE. https://publicationethics.org/topic-discussions/when-editors-suspect-ai
• COPE Council. (2025). COPE guidelines: Retraction guidelines. Committee on Publication Ethics. https://doi.org/10.24318/cope.2019.1.4
• COPE Council. (2026). Suspected AI-generated manuscripts (Case 26-11). Committee on Publication Ethics. https://publicationethics.org/guidance/case/suspected-ai-generated-manuscripts
• Elsevier. (2025, January 30). Reference checking in Editorial Manager. Elsevier Support. https://www.elsevier.support/publishing/answer/reference-checking-in-editorial-manager
• International Committee of Medical Journal Editors. (2026). Recommendations for the conduct, reporting, editing, and publication of scholarly work in medical journals (ICMJE Recommendations, updated January 2026). https://www.icmje.org/icmje-recommendations.pdf
• PLOS. (n.d.-a). Ethical publishing practice. PLOS One. https://journals.plos.org/plosone/s/ethical-publishing-practice
• PLOS. (n.d.-b). Submission guidelines. PLOS One. https://journals.plos.org/plosone/s/submission-guidelines
• Springer Nature. (2025, April 7). New research integrity AI tool added to Springer Nature's growing portfolio [Press release]. https://group.springernature.com/gp/group/media/press-releases/new-research-integrity-ai-tool/27769148
• Tang, G., & Cai, H. (2025). Citation contamination by paper mill articles in systematic reviews of the life sciences. JAMA Network Open, 8(6), Article e2515160. https://doi.org/10.1001/jamanetworkopen.2025.15160
• Taylor & Francis. (n.d.). AI policy. https://taylorandfrancis.com/our-policies/ai-policy/
• Topaz, M., Roguin, N., Gupta, P., Zhang, Z., & Peltonen, L.-M. (2026). Fabricated citations: An audit across 2·5 million biomedical papers. The Lancet, 407(10541), 1779–1781. https://doi.org/10.1016/S0140-6736(26)00603-3
• Topaz, M., Zhang, Z., Gupta, P., Roguin, N., & Peltonen, L.-M. (2026). Fabricated references are twice as common in medRxiv preprints as in peer-reviewed articles. Journal of Internal Medicine, 300(3), 379–381. https://doi.org/10.1111/joim.70139
• Zhao, Z., Wang, Y., Stuart, T., De Vaan, M., Ginsparg, P., & Yin, Y. (2026). LLM hallucinations in the wild: Large-scale evidence from non-existent citations [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2605.07723