A reference that does not exist, or that exists with another title, other authors or a DOI that leads to another work, is one of the few things that make an editor stop reading. It does not have to be fabricated by an AI: a reference manager that imported the wrong metadata, a DOI copied from the row below or a citation from memory is enough. Since 2023 editors check it, and they check it first.
Previum runs that check as part of the Basic report: it extracts the reference list from your manuscript, looks each one up in eight bibliographic databases and gives it a verdict: confirmed in two or more databases, in a single one, with the DOI pointing to another work, with a DOI no database recognizes, or with no trace. With code, not with a model: a reference is not “recalled”, it is looked up.
This page explains what is checked, how to read each verdict, what to do with each one and where the method fails (books, theses and local journals), with the figures from the public benchmark of fabricated citations we ran the verifier through. And at the end, six real, already published papers with their bibliography checked, so you can see the output before uploading anything.
Because it is the cheapest thing to check and the most expensive to let through. Verifying a DOI takes ten seconds; publishing a paper with a fabricated citation is a correction, or a retraction, with the journal’s name on it. Since language models started drafting plausible bibliographies, many journals have added a reference check to editorial screening, and some run it before assigning reviewers.
The problem is not only AI. Walters and Wilder (2023) asked two language models for 84 academic texts and checked their 636 citations by hand: 195 were fabricated. But of the 441 real ones, 60 carried some wrong detail (authors, title, year or journal). That second group is the one that most resembles what sits in an ordinary manuscript: works that exist, cited with errors a reference manager drags along for years.
And there is a third case an editor does not forgive either: the DOI that exists but leads to another article. It usually comes from copying one row too many in the spreadsheet, and it is indistinguishable from a fabrication until someone clicks the link.
When you upload the manuscript, Previum locates the reference section (the last “References” or “Bibliography” heading) and separates each entry. From each it extracts authors, year, title and DOI if there is one, and looks it up in eight databases, in this order: Crossref, OpenAlex, PubMed, Semantic Scholar, DOAJ, OpenLibrary, Google Books and arXiv. The first four cover almost everything published in journals; the last four exist for books, small open-access journals and preprints.
To count a reference as confirmed, two databases must agree on title and authors. A single database is not enough: some catalogs index whatever an author uploads by hand, and with one match you cannot tell a real citation from one someone copied wrongly in two places. That is why the “single database” verdict exists: it is not an alarm, it is a note to look yourself.
All of this is database queries and string comparison, with no language model. The model plays no part in deciding whether a reference exists: when you ask a chatbot to do that, what it does is recall, and recalling is precisely what fabricates citations.
Each reference comes out with its verdict and the databases that confirm it. The order I recommend for working through the list, from most to least serious:
These are the two that can cost you the editorial screen, because the editor sees them with one click. Open the DOI on doi.org: if it leads to another article, find the correct DOI of the work you cite (Crossref gives it in seconds from title and first author) and replace it; if it does not resolve, check it character by character against the article page, because a hyphen swapped for a full stop already breaks it. If the work has no DOI, remove it rather than leave one that does not work.
First rule out the obvious: a book, a chapter, a thesis, an agency report or an article in a regional journal can be perfectly well cited and not appear in these databases. If so, leave it alone. If it is an article in an indexed journal and it does not appear, search the literal title on Google Scholar: if it is not there either, the citation very probably comes from a draft written with a model’s help, and it must be replaced by the real source that supports that sentence, or the sentence removed.
Almost always a real citation with one wrong detail: an author too many or too few, the year of the online version instead of the print one, an abbreviated title. Open the record in the database that found it and correct the entry to match. It is the verdict that accumulates most entries in an ordinary manuscript and the least dangerous one; but a reviewer who checks three at random and finds two wrong starts to distrust the rest.
Nothing to do about the existence of the work. What the verifier does not check is that the work says what you say it says: that remains your responsibility, and it is the second thing an experienced reviewer looks at when a citation sounds forced.
A verifier that says “check your references” has to say how much it gets right, and on what. We ran ours through the 636 citations of the Walters and Wilder benchmark, with the result for each citation published: it catches 92.3% of the fabricated ones and recognizes 86.4% of the real ones; 88.2% accuracy overall, with 96.2% precision when it says a citation is real.
The 13.6% of real ones it does not recognize has an explanation worth knowing: 60 of the 441 real citations in the benchmark carry some wrong detail, and the verifier requires two databases to agree on title and authors, so it flags them. For the benchmark that counts as a miss; for you it is the warning that the citation is wrong even though the work exists. Books are the genuine weak spot: 80.9% accuracy versus 94.3% on articles, because the book databases are queried last.
Reviewer3, the tool we are most often compared with, publishes 98.5% on 476 citations it does not itemize. The figures are not comparable without the same benchmark, which is why ours come with the citation-by-citation JSON.
Evidence: the 636 citations with their label and our verdict (JSON) · Reviewer3 vs Previum, measured on its own benchmark
The check is part of the Basic report, with your account (Google, Microsoft, ORCID or an email link; any real email works). Up to 30 references are checked per pass; the full report goes to 60 and adds the audit of the ones needing action with each one’s record. Alongside go the other code checks: every p value recomputed from its degrees of freedom, the figures that do not add up (percentages against their fraction, intervals against their estimate, the n in the table against the n in the text) and APA 7 statistical reporting.
And, above all, the two substantive objections a reviewer would raise first, with the quoted sentence and the fix. A clean bibliography gets you past the editorial screen; what decides the review is the method, and the report starts there.
If you want to see the output before uploading your manuscript, there are six reports on real, already published papers, one per discipline, with their bibliography checked exactly as it came out: nursing, physiotherapy, education, sports science, dentistry and psychology.
Sample report on a real nursing paper · The six samples by discipline
Paste your reference list, up to 30 entries, and each one is looked up in Crossref, OpenAlex, PubMed, Semantic Scholar, DOAJ, OpenLibrary, Google Books and arXiv. A reference counts as confirmed when two independent sources agree on title and authorship, and confirmed references are also screened against the retraction notices in Crossref and OpenAlex. DOIs that no source recognizes, and DOIs that point to a different work, are flagged separately: the second is the most common copy error when a bibliography is reordered.
A reference that is not located is not necessarily made up: books, chapters and journals outside these databases land there. Look first at DOIs that do not match and at retracted papers. Your bibliography is not stored.
Look it up by DOI on doi.org and by literal title on Crossref or Google Scholar. If the DOI does not resolve or leads to another work, and the title appears in no database, it almost certainly comes from a draft written with a language model’s help. Previum runs that search for you in eight databases and returns each reference with its verdict, in the Basic report.
No. It means it does not appear in Crossref, OpenAlex, PubMed, Semantic Scholar, DOAJ, OpenLibrary, Google Books or arXiv. Books, chapters, theses, reports and local journals are poorly covered by those databases and can be perfectly well cited. The verdict tells you where to look; it accuses you of nothing.
Up to 30 per pass in the Basic report and up to 60 in the full one, in parallel; a list of 30 takes under a minute. If your manuscript has more, the first ones are checked and the report says so.
No. It checks that the work exists as you cite it (title, authors, year, DOI). Whether it supports the claim you cite it for remains your job; the simulated reviewer can object to a claim that does not square with what is known, but that is another part of the report.
Yes. The extractor separates entries by number, by author-year or by line, and pulls title, authors, year and DOI from each regardless of style. What it does need is the list inside the manuscript under a references heading: if you paste only the body, there is nothing to check.
No, not even with an account: it is read for the review and discarded. The database queries carry the citation, not your text. Your workspace keeps the review, which you download as Word or PDF, and it never goes out by email.