The fabricated-references crisis

A reference is the smallest unit of scientific accountability. It says: this claim does not rest on my say-so — go and check. Which is why the most quietly corrosive trend in publishing right now is the reference that cannot be checked, because the paper it points to was never written.

The numbers have moved fast. Analyses of published literature found fabricated references — citations to works that do not exist — in roughly one in 2,828 papers in 2023. By early 2026, the rate was one in 277. That is a tenfold increase in a little over two years, and it is a floor, not a ceiling: those figures count only the fabrications that automated scans and diligent readers managed to catch.

What a fabricated reference looks like

The unsettling thing about a modern fabricated reference is how unremarkable it looks. It names real researchers who plausibly could have written such a paper. It cites a real journal, a sensible year, a volume number in the right range. It often carries a DOI — correctly formatted, confidently stated, and resolving either to nothing or to an entirely unrelated article.

@article{okafor2022sleep,
  author  = {Okafor, N. and Lindqvist, H. and Tan, W.},
  title   = {Sleep restriction selectively impairs pattern
             separation in young adults},
  journal = {Journal of Cognitive Neuroscience},
  year    = {2022},
  volume  = {34},
  pages   = {1121--1138},
  doi     = {10.1162/jocn_a_01822}
}

Everything about the entry above is plausible. The authors are invented, the paper does not exist, and the DOI belongs to something else. A reader skimming a bibliography — or a reviewer with forty pages to assess and two weeks to do it — has essentially no chance of flagging it by eye. The only reliable test is the one nobody performs manually at scale: resolve every identifier and compare the returned metadata against the citation.

Where fabrications come from

Fabricated references are not new; careless or dishonest citation predates any recent technology. What changed is the mechanism of production. Large language models generate text by continuing patterns, and a citation is a pattern like any other. Asked to support a claim, a model that cannot search the literature will often produce something that is statistically shaped like a citation: real-sounding authors, a fitting title, a journal that publishes exactly this kind of work. The model is not lying in any deliberate sense. It is doing what it does everywhere — producing plausible text — in the one context where plausibility without existence is precisely the failure mode.

The rest of the pipeline is human. A draft gets assembled under deadline pressure, a generated paragraph arrives with its generated supporting citations, and the verification step — the twenty minutes of resolving DOIs and reading abstracts — is the step that gets skipped. Multiply by the adoption curve of generative drafting tools between 2023 and 2026, and the shift from one in 2,828 to one in 277 stops being mysterious. The tools got into everyone’s hands; the checking did not.

Peer review was never built to catch this

It is tempting to say peer review should catch fabricated references. In practice, it almost never does, and the reason is structural rather than a matter of diligence. Reviewers evaluate arguments, methods, and novelty. They check whether the cited literature is appropriate — not whether it exists. Existence has always been the assumption underneath the whole exercise.

The clearest demonstration came from one of the most selective venues in computer science: analyses of the NeurIPS 2025 proceedings identified more than 100 hallucinated citations in papers that passed review. These manuscripts cleared multiple expert reviewers and an area chair. If fabricated references survive that process at a venue with single-digit-to-low-twenties acceptance rates, no journal can reasonably promise its review pipeline filters them out.

Reviewers check whether cited literature is appropriate — not whether it exists. Existence was always the assumption. That assumption no longer holds.

Why it matters more than it seems

A single phantom citation might look like a victimless typo. It is not, for three reasons. First, references are load-bearing: a claim cited to a nonexistent source is an unsupported claim wearing a disguise, and readers who trust the disguise propagate the claim onward. Second, fabrications poison the citation graph. Bibliometric databases, literature-mapping tools, and systematic reviews all assume the graph’s nodes are real; every phantom entry quietly corrupts the systems the research community uses to navigate itself. Third, the reputational asymmetry is brutal. For the author, a discovered fabrication reads as misconduct regardless of intent — “the tool made it up” is not a defense any integrity committee accepts, because accountability for the reference list never left the author.

There is also a collective cost that compounds: suspicion. As detection tooling improves and more cases surface, every bibliography inherits a small tax of doubt. The one-in-277 figure means a working researcher now encounters fabricated references routinely, whether they notice or not — and editors are beginning to treat reference verification as a screening step rather than an assumption.

What verification-first tooling actually means

The response to this crisis is not to ban assistance; that ship has sailed, and assistance is genuinely useful. The response is to change where trust sits in the pipeline. A verification-first writing environment inverts the default: instead of trusting citations until someone disproves them, no reference enters the manuscript without being resolved against the scholarly record first.

Concretely, that means a few enforceable properties:

• Resolution at insertion. Every citation added to the text is resolved against a registry such as Crossref at the moment it is added — by DOI, with returned metadata compared to the claimed title, authors, and venue. An entry that does not resolve is visibly flagged and cannot silently join the reference list.
• Search, never generate. If an assistant proposes supporting literature, the proposals must come from querying real, resolvable records — with the metadata attached — rather than from generating citation-shaped text. An assistant that cannot find a source should say so, because a confident fabrication is far more expensive than an honest blank.
• Continuous re-checking. References rot even when honest: DOIs die, preprints acquire published versions, entries duplicate. The reference list should be re-validated as the manuscript evolves, with repairs proposed as reviewable diffs.
• Claim–citation support. Existence is the floor, not the goal. The harder question — does the cited source actually support the sentence citing it? — is increasingly checkable too, and pre-submission checks should raise it while the fix is still cheap.

This is the posture we build into Dissertatio: citations resolve against Crossref when they enter the manuscript, the assistant can only cite records it has actually retrieved, and validation re-runs on every edit so a reference list that was clean on Tuesday cannot quietly decay by Friday. But the principle matters more than any product: verification belongs in the writing environment, not in the reviewer’s spare time.

What authors and labs can do this week

Even without new tooling, the exposure is manageable with a little process:

• Resolve every DOI in the bibliography before submission, and check the returned title and authors — not just that the link opens.
• Treat any generated text that arrives with citations as containing unverified claims by default, and verify before it touches the draft.
• Make reference verification an explicit, named step in the lab’s submission checklist, owned by a specific author — unowned steps are skipped steps.
• Keep records of how assistance was used in drafting, so disclosure statements can be written from logs rather than memory.

The fabricated-references crisis is, at root, a trust-allocation bug: tools that generate plausibility were plugged into a system that runs on verified existence. The fix is not less tooling but better-aimed tooling — and a research culture that treats “every reference resolves” the way software teams treat a passing build: not an aspiration, but the precondition for shipping.