The Hallucinated-Citation Audit: Pre-Submission Checks

Maxim Topaz, a Columbia team that audited 2.5 million biomedical papers indexed in PubMed Central, has uncovered what we all secretly feared. The number of fake references being included into peer-reviewed biomedical print increased sharply, precisely when generative AI tools became part of the average researcher's drafting workflow. One out of every 277 papers cited in the first seven weeks of 2026 cites at least one reference that isn't to be found anywhere in PubMed, Crossref, OpenAlex, or Google Scholar. Two years earlier, the proportion was one in 2,828. We're looking at a 12-fold increase in the rate of fake references finding their way into the literature, with the greatest uptick starting right on schedule in mid-2024.

The Topaz audit is the cleanest number we have, but it isn't the only one. ArXiv announced a one-year submission ban for authors who post papers with unambiguous AI-generated errors, including hallucinated citations. GPTZero's analysis of 4,841 NeurIPS 2025 accepted papers identified at least 100 confirmed hallucinated citations in 53 papers after they had been vetted by three to five expert reviewers. A Deakin University study of GPT-4o-generated references found that 56 percent were either fake or contained errors, and 64 percent of the fake ones came packaged with a DOI that linked to a real but unrelated paper, the variety that's hardest to catch.

The good news is that a pre-submission hallucinated citation audit takes between 20 minutes and two hours depending on the size of your reference list, and it catches almost all of these errors before an editor or a reviewer sees them. This post is the audit. We cover what hallucinated citations actually look like in 2026, the four-database verification workflow, the step-by-step pre-submission checklist, the tools that automate the work, and what to do if you find one in a manuscript that is already under review.

Why hallucinated citations are now a screening problem

In about the second half of 2024, editors start noticing hallucinated references appearing on their technical screen. In mid-2026, all leading medical and life-sciences journals include a reference-existence check into their in-house pre-review procedures or guidance for peer reviewers. The Lancet letter that published the Topaz audit was, in part, a public commitment to harden the journal's own intake.

The reason this matters operationally is that a hallucinated citation is no longer treated as a careless error. It is treated as a research integrity flag. The leap from "the authors got a year wrong" to "the authors did not read this source because it does not exist" is short, and editors are now trained to make it. At most journals, one made reference in a 60-entry list is enough to bounce the manuscript; at the strictest ones (the Lancet, BMJ, several Nature Portfolio titles), it can trigger a referral to the institution's research integrity office.

The arXiv policy makes the same point in sharper form. Any author submitting a paper containing a hallucinated reference may be banned from arXiv for a year. After the ban every next arXiv submission must go through peer review at a reputable venue before arXiv will accept it. This is a permanent reputational asymmetry based on a single unchecked citation. The fake citation pubmed problem isn't just a PubMed problem.

What hallucinated citations actually look like

Most authors imagine a hallucinated citation as obviously broken: a malformed title, a journal name that does not exist, a year set in the future. Those exist, and they are the easiest 30 percent of the problem. The harder 70 percent looks legitimate and survives a quick visual check.

We see four recurring patterns.

Pattern 1: real authors, real journal, plausible title, no such paper. The model assembled the citation from training data that taught it which authors publish where and on what topics. The output reads like a real reference. The DOI, if present, either does not resolve or resolves to a different paper.

Pattern 2: real paper, wrong author or wrong year. The model knew the paper existed but mismatched a metadata field. The citation passes a Google Scholar title search but fails a DOI lookup. This is often the hardest to spot, since the cited claim might actually be present in the real paper (under a different author or year).

Pattern 3: real DOI, unrelated paper. The Deakin study finding. The model produced a plausible-looking citation and attached a DOI that does resolve, but to a completely unrelated piece of work. A visual check illustrates the DOI is real, but only reading the linked paper illustrates the problem. This is the 64 percent variety.

Pattern 4: tortured-phrase chains around the citation. The Cabanac Problematic Paper Screener has flagged more than 19,000 papers containing tortured phrases generated by paraphrasing software, often clustered around made citations. "Joined Together States" for United States, "bosom peril" for breast cancer, "kidney disappointment" for kidney failure. If the prose around a citation reads as paraphrase-laundered, the citation is more likely to be hallucinated too. Our note on paraphrasing that preserves citations covers the technique gap that produces Pattern 4 in the first place. Pattern 4 is correlated with Patterns 1 to 3, not independent of them.

These patterns are the reason a single-database check is not enough. A title-only Google Scholar lookup misses Patterns 2 and 3. A DOI-only Crossref lookup misses Pattern 1. The four-database workflow exists because each database catches a different failure mode.

The four-database verification workflow

We define a hallucinated reference as one that isn't present in any of PubMed, Crossref, OpenAlex, or Google Scholar. For our pre-submission audits, we rely on these four sources because they each cover holes in the other three.

PubMed. Best coverage for biomedical and clinical literature. If one's reference has a PMID, use the PMID lookup. Otherwise, look up the reference by title and first author's last name. PubMed will reject many real-but-wrong citations that Crossref accepts.

Crossref. Best coverage of DOIs across all fields. Look up the reference by DOI using this doi.crossref.org lookup or via their public API at api.crossref.org/works/[DOI]. If it doesn't work, then it's definitely hallucinated. If it does, check out the title, authors, and year.

OpenAlex. The widest coverage, free, and it uncovers citations that aren't covered by PubMed and Crossref (humanities, computer-science conference proceedings, regional journals). Look up the reference by title and make sure the authors are right. OpenAlex is also useful because it surfaces citations that exist as preprints but not yet as DOIed publications.

Google Scholar. Use last, as the broadest-coverage but lowest-precision check. A reference that doesn't appear in any of PubMed, Crossref, OpenAlex, or Google Scholar is functionally certainly hallucinated. A reference that appears only in Google Scholar but in none of the structured databases is suspect: it may exist (gray literature, a thesis, a non-DOIed proceedings) but the citation needs a manual check before submission.

The workflow takes roughly one minute per reference if done manually. A 60-reference paper is an hour of work. A 200-reference systematic review is closer to three. The audit is tedious but mechanical, which makes it a natural target for automation; tools are in the next-but-one section.

The pre-submission audit, step by step

A repeatable checklist we use on every manuscript that came near a generative AI tool during drafting. Eight steps, none of which requires a paid tool.

1. Extract the reference list to a flat format. Export from your reference manager (Zotero, Mendeley, EndNote, Paperpile) as BibTeX or RIS. If your paper was drafted in Word without a manager, paste the reference list into a plain text file. Number each entry. You will need the numbering to track which references have been verified.

2. Flag any reference whose title you do not recognize. If one cannot picture the paper, one almost certainly did not read it. Hallucinated references are at the highest risk of being false. For the most thorough check, mark them.

3. Resolve every DOI. Paste each DOI into https://doi.org/[DOI]. If the DOI does not resolve, the reference is hallucinated. If it resolves, compare the destination paper's title and authors against one's citation. Mismatches are Pattern 3 hallucinations.

4. Verify every non-DOI reference in PubMed or Crossref. Title search plus first-author surname. If the paper exists, capture the PMID or DOI for one's records. If it does not exist in either database, escalate to step 5.

5. Cross-check escalated references in OpenAlex and Google Scholar. If a reference fails Crossref and PubMed, the OpenAlex and Google Scholar checks are the deciding lookup. Failure in all four databases is one's evidence that the reference is hallucinated. If a reference appears only in Google Scholar, manually inspect the linked source.

6. For preprints, verify both the preprint and any published version. A common mid-2024-and-later pattern: the model cites a real preprint but invents a fictional journal publication of the same work. Check arXiv, bioRxiv, medRxiv, and SSRN explicitly for any preprint reference.

7. Spot-check the in-text claim against the source. Open the actual paper and check that it supports the specific claim you are citing it for in at least 10 percent of one's references (or all of them, for the systematic-review case). The Deakin study's 56 percent error rate includes Pattern 2 cases where the cited paper exists but does not support the claim.

8. Document the audit. Save the audit log (which references one verified, against which databases, on what date). If the journal asks during peer review, one has receipts. If a reviewer challenges a specific citation, one can point to one's verification step.

The cost of running the audit is two hours. The cost of skipping it, if one gets caught, ranges from technical-screen rejection to the arXiv ban. Run the audit.

Tools that automate the audit

You can do all eight steps manually. You'll be faster with tooling.

Crossref public API. Free, no signup needed. The endpoint api.crossref.org/works/[DOI] returns JSON with the canonical title, authors, and year. A 20-line Python script that walks your reference list and hits this endpoint catches Patterns 1 and 3 in seconds per reference. This is the lightest possible automation and is usually enough for a 60-reference paper.

Scite (scite.ai). Verifies references and reports on whether next papers cited the work supportingly or contradictorily. Catches Pattern 2 (real paper, wrong claim) better than the other tools because it engages with the actual content. Has a free tier; the paid version is necessary for full integration.

Problematic Paper Screener (Cabanac, Labbe, Magazinov). Free, open, largely a post-publication screener but its detector suite is useful pre-submission for Pattern 4. Search at dbrech.irit.fr/pls/apex/f?p=9999:1. Useful for confirming that the prose around your citations hasn't been paraphrase-laundered to the point where it looks like paper-mill output.

Reference checkers built into reference managers. If Zotero sees a hallucinated reference, its "Find Available PDF" function will not work. Mendeley's feature to fill in metadata has the same effect. This is not a purposeful hallucination checker, but it does identify the easiest problems when you're using it normally.

Our proofreader. Our AI proofreader works by running the four-database audit in one pass, highlighting hallucinated and mismatched references, and exporting a clean log you can save with the manuscript. We provide this tool as part of our wider suite of AI proofreading tools for research papers, covering the associated grammar, citation-formatting, and tracked-changes processes that come with a clean reference audit.

Most of the time, the Crossref API script and a manual escalation list will catch 90 percent of the problem for most authors. Tools earn their value for systematic reviews, for research groups with many submissions, and for the case where the cost of a single missed hallucination is big (a Lancet or NEJM submission, an arXiv repost after a previous flag).

What to do if you find a hallucinated citation

You will find one. In our experience, almost every reference list that touched an LLM has at least one. Depending on your manuscript's stage, you have two options.

Pre-submission. Easiest case. Remove the citation, or replace it with a real reference that actually supports the claim, or rewrite the surrounding sentence so the unsupported claim is removed. Document the change in one's audit log.

Under peer review. Notify the handling editor in writing as soon as one confirms the hallucination. Provide the corrected reference list and an explanation: what one found, how one found it, what one has done to verify the remaining references. Editors respond much better to authors who self-correct than to authors caught by a reviewer.

Accepted but not yet published. Contact the production editor right away. Most publishers will accept a late correction at the proofs stage if it's presented as integrity self-correction rather than a routine edit.

Published. Contact the journal and request an erratum or, in serious cases, a corrigendum. If the hallucinated citation supports a substantive claim in the paper, the correction may need to extend to the claim itself. This is the path one most wants to avoid; it is the path the audit prevents.

AI disclosure statement on one's manuscript if one used any AI tools that touched the reference list, such as paraphrasers or citation generators. One can also disclose in the cover letter that one ran an audit and self-corrected. It sounds like a professional thing to do rather than alarming.

Reference-Checking Proofreader for Journal Submissions

One-pass hallucinated-citation audit against PubMed, Crossref, OpenAlex, and Google Scholar, with a downloadable verification log for your records.

Frequently asked questions

Q: How common are hallucinated citations in 2026?

The Columbia analysis (Topaz et al, Lancet 2026) found that one in 277 PubMed-indexed papers published in the first seven weeks of 2026 contained at least one reference. That's a 12-fold increase over the 2023 rate of one in 2,828. The trend is going up. At conferences with strict peer review, the GPTZero NeurIPS 2025 analysis found at least 100 hallucinated citations in 53 of the 4,841 accepted papers, around 1 percent of the conference.

Q: Why is the 64 percent DOI variety the hardest to catch?

When an LLM hallucinates a reference, it sometimes puts a real DOI from its training data that points to another paper than what was cited. The DOI resolves, the other paper exists, and a quick visual check seems OK. It's only after clicking on the link to the referenced paper that one realizes that the title, authors, or even contents of the cited reference do not match the citation. The Deakin study illustrates that this was true for 64 percent of the fake citations generated with GPT-4o. Just resolving the DOI is insufficient and needs to be complemented by verification that the linked paper indeed corresponds to the citation.

Q: Will arXiv really ban me for one year for a hallucinated citation?

Yes, when the negligence is unambiguous. Thomas Dietterich announces the May 2026 arXiv policy that aims to address cases where authors pasted LLM output without checking it, which led to hallucinated references, leftover chatbot placeholder text, or data. ArXiv has been explicit that truly contested or borderline cases are out of scope. After the ban expires, all future arXiv submissions must first clear peer review at a reputable venue before arXiv will re-accept them.

Q: Can I just run my reference list through ChatGPT or Claude to check it?

No. The same model that hallucinated the reference can also hallucinate the verification. We have seen LLMs confidently "confirm" that made DOIs are valid. The check has to be against external structured databases (PubMed, Crossref, OpenAlex, Google Scholar) that have authoritative metadata, not against another LLM.

Q: What if a reference exists only on Google Scholar?

A reference that appears in Google Scholar but in none of PubMed, Crossref, or OpenAlex is suspect rather than confirmed-hallucinated. Some real sources (a thesis, a conference proceedings without a DOI, a working paper, a regional journal) appear in Google Scholar without structured-database coverage. Manually inspect the linked source. If one can open the paper and verify the claim, the reference is real but under-indexed; cite it with confidence. If one cannot open the source or the link is broken, treat it as hallucinated and remove or replace.