How to Check Whether a Submission’s References Are Fabricated
Fabricated references are no longer a rare curiosity in scholarly submissions. They are a routine screening problem, and most small journals have no process for catching them. A reference that does not exist reads exactly like one that does: plausible authors, a real journal, a volume and issue that look right, page numbers in a sensible range. Nothing about it looks wrong until someone tries to find it — and by then it is often printed.
This article sets out what fabricated references actually look like in a submitted manuscript, why they have become common, the four checks that catch almost all of them, and how to raise the problem with an author without accusing anyone of misconduct.
Why fabricated references suddenly matter
Three things changed at once.
First, large language models write fluent citations that do not exist. When a model is asked to support a claim, it produces a reference shaped like a real one — because it has learned the shape, not the record. Authors who draft with these tools and do not verify the output submit those citations intact. This is usually carelessness rather than fraud, but the result on your pages is identical.
Second, reference managers made bulk importing easy and verification invisible. A citation dragged from a search result into a manuscript carries whatever metadata was attached to it, including metadata that was wrong at the point of capture.
Third, paper mills sell manuscripts with padded reference lists. Citations are added to make a paper look embedded in a literature it never engaged with, and to inflate the citation counts of cooperating authors.
The consequence for an editor is that reference lists can no longer be assumed correct because they look correct.
What a fabricated reference looks like
In practice, what arrives in your inbox falls into four categories, and only the first is a true fabrication.
1. The reference does not exist at all
No record in Crossref, PubMed or OpenAlex. Frequently the journal is real and the volume is real, but that article was never published. A common signature: the page range belongs to a different paper in the same issue. Another: a DOI that is correctly formatted and resolves to nothing, or resolves to something unrelated.
2. The reference exists but does not match
Far more common than outright fabrication. The article is real, but the author list is wrong, the year is off by one, the journal name has been swapped for a similar title, or the title has drifted. These are usually honest errors — but a mismatched author list is also exactly what a fabricated citation looks like after someone half-corrected it.
3. The reference has been retracted
The paper exists, was published, and has since been retracted or corrected — and the manuscript cites it as current evidence with no acknowledgement. This is the category most likely to embarrass a journal after publication, because the retraction is a matter of public record and anyone can find it.
4. The reference is real but does not support the claim
The hardest category, and outside the reach of any automated check. The citation resolves perfectly. It simply does not say what the sentence attached to it claims. Catching this requires reading the cited paper, which is why it is normally done as a sample rather than exhaustively.
The four checks that catch almost everything
Check one: does the record exist?
Resolve every reference against Crossref, PubMed and OpenAlex. Use all three, not one. Crossref covers anything with a registered DOI, PubMed covers biomedical literature including older material without DOIs, and OpenAlex catches a great deal that neither holds — conference proceedings, regional journals, preprints. A reference absent from all three is not automatically fabricated, but it needs a human to look.
Do this by DOI where one is given. Where none is given, match on title plus first author plus year, and accept fuzzy title matching — real citations contain typographical errors constantly.
Check two: does every field match?
Existence is not enough. Compare the submitted authors, year, journal, volume, issue and pages against the retrieved record, field by field. Flag any disagreement. Most disagreements are trivial and get corrected by the author in five minutes; the value of the check is that the small proportion which are not trivial become visible.
Pay particular attention to author-list mismatches. A wrong page number is a typing error. A wrong second author on a paper cited as central evidence is worth a question.
Check three: has anything been retracted?
Crossref carries retraction and correction assertions in its metadata, and the Retraction Watch database records a great deal more. Screen every reference against retraction status, not just the ones that look suspect. Retracted papers continue to be cited for years after retraction, usually because the citing author copied the reference from another paper’s bibliography and never opened it.
When you find one, the question to the author is simple and not accusatory: are you aware this has been retracted, and does your argument survive without it?
Check four: what shape is the list?
Two summary statistics are worth producing for every manuscript. The first is self-citation rate — what proportion of the list is the authors citing themselves. There is no correct number, and legitimate reasons for a high rate exist, but a list where a third of the references share an author with the submission deserves a look. The second is source concentration: whether one journal, one research group or one review article is doing an unusual amount of the work. Padding is visible in the shape of a list before it is visible in any individual line.
What to do with the flags
A flag is not an accusation, and how the finding is communicated determines whether you get a corrected manuscript or a defensive author.
The overwhelming majority of reference problems are errors. Authors mistype, reference managers corrupt entries, and co-authors inherit bibliographies from earlier drafts. Treat the default explanation as error, and write to the author accordingly: here are the references that could not be verified, here is what our check found, please correct or supply the correct source.
Escalate only on pattern. One non-existent reference is an error. Several non-existent references, concentrated in the passages that carry the paper’s central claims, is a different matter — and at that point the relevant framework is the COPE guidance on suspected misconduct, handled as a case rather than a correction. The ICMJE Recommendations set out the author responsibilities you can reasonably point to.
Do not put an accusation in writing on the strength of an automated flag. Ask a question instead.
When in the workflow this belongs
Reference checking belongs before peer review, not after acceptance. Three reasons.
It saves reviewer goodwill. Reviewers who receive papers with broken evidence bases stop answering your invitations, and reviewer supply is the scarcest resource a small journal has.
It is cheapest at the point where the paper might still be rejected. Running a deep check on a manuscript you are going to desk reject anyway wastes the check.
And it produces a document. If an author later disputes a correction or a rejection, a dated report showing exactly what was checked and what was found is far stronger than a recollection.
The practical sequence for most journals: a completeness and scope screen first, a reference check on everything that survives it, and the deeper checks — figures, claim substantiation — reserved for papers heading towards acceptance.
Doing it yourself, or having it done
The mechanics above are reproducible by hand. For a list of forty references, expect an experienced person to take between one and two hours to do all four checks properly, most of it spent on the ones that fail first-pass matching. For a journal receiving ten submissions a month, that is a part-time job nobody currently has.
The alternative is to have it run for you. NeucitePress checks every reference in a submitted manuscript against Crossref, PubMed and OpenAlex, flags non-existent, mismatched and retracted entries, reports self-citation and source concentration, and returns a numbered correction list addressed to the author — in one working day, at $18 per paper. There is no contract and no minimum; most journals start by sending the one paper they are least sure about.
It sits alongside the rest of our submission screening services: a scored desk-decision report with the rejection letter already drafted, figure integrity checks, author and reviewer identity verification, and a sampled citation substantiation audit for the claims that matter most. Journals that want the whole desk covered use our managed editorial office, and those building the governance around it start with the ethics and governance programme.
Frequently asked questions
Can software prove a reference was fabricated?
No. Software can prove a reference cannot be found in the major indexes, which is not the same thing. Legitimate work exists outside Crossref, PubMed and OpenAlex — regional journals, older material, books, grey literature, government reports. A “not found” result means a person needs to look, not that the author invented it.
How many fabricated references are normal?
There is no established baseline, and any figure quoted as one should be treated sceptically. What is worth tracking is your own journal’s rate over time, because a change in it tells you something even when the absolute number does not.
Should we check references before or after peer review?
Before. Reviewers should not spend their time on a paper whose evidence base has not been verified, and a reference problem found before review can be fixed by the author without a second review round.
What do we tell an author whose reference does not exist?
Ask, do not allege. State which reference could not be verified, say what was searched, and ask them to supply the correct source or remove the claim. Most respond with a corrected citation within a day.
Does a retracted reference mean the manuscript must be rejected?
Not by itself. It means the author must acknowledge the retraction and show whether the argument survives without that source. Whether it does is an editorial judgement, and it stays with your editors.