Hallucinated Citations Are a Sixfold Problem. Here's How to Catch Them.

A study published in the Lancet on May 7, 2026 puts a number on something many editors have suspected for two years. Fabricated citations in academic literature are climbing fast, and the curve is steep enough that any author submitting in 2026 needs a verification step they did not need in 2022.

1 in 277 papers. That is the rate of fabricated references in the first seven weeks of 2026, up from 1 in 2,828 in 2023 — a sixfold increase in two years, still accelerating.

The Columbia team analyzed roughly two million papers and 97 million citations. In 2023, fabricated references appeared in 1 in 2,828 papers. By 2025 the rate was 1 in 458. In the first seven weeks of 2026, it was 1 in 277. The analysis identified about 4,000 fabricated citations across 2,800 papers. Over a third of them came from two large open-access publishers (StatNews coverage).

The mechanism is well understood at this point. Generative models, asked to support a claim, will produce plausible references to papers that do not exist. The author names are real. The journal is real. The year is in range. The DOI sometimes resolves. The paper itself was never written.

What the Lancet study clarifies is that this is no longer a niche problem confined to obvious AI-written submissions. It is showing up in manuscripts that pass internal review, that get past co-authors, and that reach published versions of record at indexed journals.

#Why Fabricated Citations Slip Through

Three failure modes overlap in the typical case.

The first is the way authors are using the tools. A researcher pastes a draft into a chatbot and asks for "supporting references for this paragraph." The model produces a list. The author skims, sees plausible names and journals, and pastes the citations into the manuscript. There is no verification step because the output looks like the kind of list a careful colleague might have produced.

The second is that bibliographies are read differently from body text. Co-authors and reviewers scan the references for missing classics and obvious errors, not for whether each entry corresponds to a real paper. A reference that looks correctly formatted, with a believable author and a recent year, gets the same eye-pass as a real one.

The third is the cultural shift Mohammad Hosseini at Northwestern named for the Lancet study: citation as box-checking. When a citation's purpose is to satisfy a reviewer comment or hit an introduction's expected density, the act of citing has been decoupled from the act of reading. A model-generated reference fills the slot just as well as a real one until somebody checks.

#Why Traditional Detection Fails

Several major publishers have rolled out reference-checking tools, and the Lancet study notes that these tools struggle with realistic inputs. Two specific failure modes appear repeatedly.

Formatting heterogeneity. The same reference can appear with the title quoted, abbreviated journal names, abbreviated author lists, missing volume numbers, or an in-press notation. A regex-based or strict-parser approach drops a high percentage of real references as unparseable, which buries the fabricated ones in noise.

The "real DOI, wrong paper" pattern. A model sometimes supplies a real-looking DOI that resolves to a paper unrelated to the claim it is supporting. A naive checker confirms the DOI exists, marks the citation as verified, and moves on. The citation is wrong in a way that only a content comparison catches.

A useful check has to handle messy formatting, query multiple authoritative databases, and compare what the database returns against what the manuscript claims, not just whether something resolves.

If you are about to submit, run a citation pass on your bibliography before you do. Most authors find one or two entries that need attention.

#What Automated Verification Actually Does

ManuscriptMind's citation check runs as a separate agent in the review pipeline. The pieces are deliberately conservative.

Reference extraction. The bibliography is parsed out of the manuscript and each entry is converted into structured fields: title, authors, year, DOI, venue. The raw text is preserved alongside the parsed fields so the report can quote the original.

Crossref DOI resolution. Where a DOI is present, it is resolved against Crossref. If the DOI resolves to a paper whose title does not match the citation's title, the entry is flagged as hallucinated rather than verified. This catches the "real DOI, wrong paper" failure mode directly.

Title-based lookup against Crossref and Semantic Scholar. Where no DOI is given, the title and authors are queried against both databases. The best candidate is scored by normalized title similarity. A high match is verified automatically. A low match drops to the next stage.

LLM judgment for ambiguous cases. When the database returns a possible match that is close but not conclusive, the candidate set is handed to an evaluator that returns one of four statuses: verified, partial_match, hallucinated, or unverifiable. The fourth category is important and often missing from naive tools. A 2025 conference paper that has not yet been indexed should not be called fabricated; it should be called unverifiable, and the author told to confirm it manually.

The output is a report listing every reference in the bibliography with its status, the matched paper where one exists, and a note explaining the judgment. Hallucinated and partial_match entries get surfaced first.

#What This Means for Authors

The practical change is small and worth doing every time you submit.

Run a verification pass after the manuscript is otherwise complete and before it leaves your hands. The tool will flag a handful of entries that need attention. Most of those will be real papers with metadata errors, which are worth fixing on their own. A smaller number will be entries that no database recognizes, and those are the ones that decide whether you submit a clean manuscript or one with a fabricated reference.

Two specific habits help.

Never paste a model-generated citation list into a manuscript without checking each entry. This is the single behavior change that prevents the most common failure. If the model gave you a reference, look it up before you cite it.

Treat any unverifiable entry as a question, not an answer. If a verifier cannot find a paper, that does not always mean the paper is fake. It does always mean you should be able to point to where you read it, what it said, and why you cited it.

The Lancet study's framing is uncomfortable but accurate. The rise in fabricated citations is partly a tool problem and partly a culture problem, and the culture problem is older than the tools. The tool problem is the one that has a fix available right now.

#How ManuscriptMind Helps

Upload a manuscript and the citation agent runs alongside the standard review. You get a per-reference report that lists every entry in your bibliography with one of four statuses — verified, partial match, hallucinated, or unverifiable — the matched paper where one exists, and a one-line note on the judgment. Hallucinated and partial-match entries surface first. Each lookup goes against Crossref and Semantic Scholar; LLM judgment is reserved for the ambiguous middle cases databases alone cannot decide. Most bibliographies finish in under two minutes.

Catch fabricated citations before submission →

---

Related: 5 Common Methodology Issues That Get Manuscripts Rejected and How to Respond to Reviewer Comments Without Tanking Your Resubmission.