The Ghost References Crisis: How AI Hallucinations Are Contaminating Scientific Literature
Picture this: a librarian at a research institution is reviewing a newly published Springer Nature paper on bowel surgery management. She checks the references one by one. Of the 14 citations listed, 12 lead nowhere. The papers do not exist. They never did.
This is not a cautionary hypothetical. It was reported by Retraction Watch in March 2026, and it is one of dozens of similar cases surfacing across the academic literature right now.
The Numbers Tell a Troubling Story
In May 2026, a landmark study published in The Lancet by Columbia University researcher Maxim Topaz and colleagues put hard figures to what many in academia had been sensing for months. After auditing nearly 2.5 million PubMed-indexed papers and verifying 97.1 million references, the team identified over 4,400 fabricated citations across 2,810 published articles. [1]
The raw count is alarming. The trend is worse. Fabricated citations in the biomedical literature have increased 12-fold in two years. In 2023, 1 in 2,828 papers contained one or more fabricated references. By 2025 that figure had reached 1 in 458, a sixfold increase. During the first seven weeks of 2026, the rate reached 1 in 277 papers. [1], [2]
The sharpest increase coincided with the rise of AI writing tools beginning in mid-2024. That timing is not coincidental. It is a direct consequence of how large language models are being used across academic writing workflows.
Critically, over 98% of the articles found to have fake references had seen no publisher action at the time of the audit. The contamination is already in the record. [2]
The sharpest increase coincided with the rise of AI writing tools beginning in mid-2024. That timing is not coincidental. It is a direct consequence of how large language models are being used across academic writing workflows.
Critically, over 98% of the articles found to have fake references had seen no publisher action at the time of the audit. The contamination is already in the record. [2]
A Researcher's Wake-Up Call
The new analysis was spurred by an unfortunate personal experience. Maxim Topaz used an AI chatbot to help edit an editorial he was submitting to a journal. Even after checking to make sure the citations were accurate, after several stages of edits, an editor at the journal flagged an erroneous citation.
"I was deeply embarrassed: I checked for that, and it still almost happened to me," Topaz said. [2] That experience drove him to investigate how widespread the problem had become across the broader literature. If it could happen to a researcher actively studying AI in healthcare, who was specifically looking for errors, it could happen to anyone.
What Is AI Hallucination, and Why Does It Affect References?
AI hallucination refers to the tendency of language models to generate text that is fluent and confident, but factually wrong. The model has no concept of truth. What it has is a statistical understanding of how text patterns fit together. When asked to generate a citation, it produces something that looks correct: a convincing author name, a realistic journal title, a credible paper title, a year that fits. The problem is that this construction may correspond to nothing that exists in the literature.
Several factors allow these fabrications to slip through unnoticed, and understanding them makes clear that this is a systems problem, not a personal failing.
- Language models are not connected to bibliographic databases. They learn what references look like, not what references are. When a researcher asks an AI tool to suggest citations, the model draws on statistical associations between topics, author names, and journal titles, rather than querying Crossref or PubMed. The output can look entirely convincing even when it corresponds to nothing real. [3]
- Researchers are working under enormous pressure. Academic publishing demands are relentless. Deadlines are real, teaching loads are heavy, and grant timelines do not pause for thorough reference audits. When a researcher turns to an AI tool to help manage that workload, the reasonable assumption is that the output can be trusted. That assumption, unfortunately, is not one these tools have earned yet. [2]
- AI tools have made the problem genuinely hard to spot. The references hallucinated by language models are not obviously wrong. They carry recognisable author names, legitimate-sounding journal titles, and believable publication years. Even a careful reader reviewing the list would have little reason to suspect anything was amiss. This is not a failure of diligence so much as a failure of these tools to signal their own unreliability.[3]
Three Forms Fabricated References Take
Not all hallucinated references look alike. They tend to appear in three recognizable patterns.
Fully Invented Citations
The most visible form is a reference invented from scratch: a credible-sounding title, a real-sounding author, a legitimate journal name, and volume and page numbers that fit the format. Everything looks right on the surface. Nothing resolves in a database.
One paper flagged in the Lancet study had 18 of its 30 references appear to be fabricated. More than half of its entire citation apparatus was fiction, yet it had been published and indexed without triggering any alarm. [1]
The “Patchwork” Reference
A more subtle variant involves real components assembled into something that never existed: a genuine researcher's name, a real journal, a well-constructed title, and a DOI that closely resembles a real one but leads nowhere or to a completely different paper. These references exploit the trust readers place in recognizable elements. Each piece looks legitimate. The fabrication is in how they are assembled.
A striking example of this came to light via Nature. [3] Computer scientist Guillaume Cabanac, who researches fabricated papers at the University of Toulouse, received a Google Scholar alert that his work had been cited in the International Dental Journal, a field entirely unrelated to his research. When he looked up the reference, he did not recognise his own supposed publication. His name had been attached to a paper he never wrote.
Real Papers, Wrong Claims
The hardest form to detect involves real papers cited in support of claims they do not actually make. The reference resolves correctly. The paper exists. It simply does not say what the citing author claims it says. Detecting this requires reading the cited work, not just verifying that it exists, and it is what one integrity researcher described as the deeper and more troubling problem lurking beneath the headline statistics. [3]
Why the Stakes Are Higher Than They Appear
Review articles had a fabrication rate 57% higher than other paper types, which is precisely the category most relied upon for clinical guidelines and evidence synthesis. [1] A systematic review built on papers with hallucinated references carries that uncertainty forward into medical practice.
The problem also extends beyond academia. Retraction Watch reported that a World Bank paper on obesity trends contained at least 14 fake references. [4] When fabricated citations appear in policy documents, decisions affecting public health and public spending may rest on sources that simply do not exist.
How to Protect Your Work
The problem is significant, but it is not unmanageable. A few straightforward habits, built into your existing workflow, can protect your work and your reputation.
- Treat AI-generated references as drafts, not finished citations. AI writing tools are not connected to bibliographic databases. Every reference they produce should be independently verified before submission.
- Verify DOIs directly. A hallucinated DOI follows the structural pattern of a real one but resolves to nothing, or to a different paper entirely. Paste each DOI into doi.org and confirm it lands where the reference claims.
- Use reference management software. Tools like Zotero, Mendeley, and EndNote pull metadata from real databases. If a paper does not exist, the search returns nothing, which is exactly the signal you need.
- Use CiteOrbit to cross-check your full reference list. CiteOrbit was built specifically for this problem. When you upload your reference list, CiteOrbit cross-checks every citation against major academic databases, including Crossref, and Google Scholar, and flags any reference that cannot be verified. But verification is only part of what it does. CiteOrbit also checks your references for formatting inconsistencies, incomplete metadata, mismatched citation styles, and other structural errors that can hold up peer review or result in correction requests after publication. Suspicious entries are highlighted before your paper ever reaches an editor, and formatting issues are flagged alongside them, so you can submit with confidence on both fronts. A reference list that might take hours to audit manually takes seconds with CiteOrbit.
- Disclose your AI use where required. Many journals now ask authors to declare whether AI tools were used in preparing a manuscript. Treating this as a routine part of submission, rather than a burden, helps the broader research community build an accurate picture of where the technology is being used and where guidance is still needed.
A Shared Problem That Needs Shared Solutions
No researcher sets out to publish fabricated references. The people affected by this crisis are, for the most part, working hard under difficult conditions and trusting tools that present themselves as reliable. The problem is structural, and addressing it requires structural responses: better tooling, clearer guidance from journals, and a shared understanding of where AI writing assistance currently falls short.
Hallucinating is inextricably linked to how large language models operate. This is not a bug awaiting a patch. It is a feature of the technology that makes verification a necessary part of any AI-assisted writing workflow, not an optional extra.
The good news is that verification is no longer as burdensome as it once was. The right tools make it fast, systematic, and genuinely reassuring. The references that are already in the literature cannot be recalled. But the ones in your next paper can be checked before anyone else ever sees them.
CiteOrbit helps researchers, authors, and institutions verify academic references against major global databases, flagging suspicious or unverifiable citations before submission. Learn more at citeorbit.com.
References
[1] Topaz M, Roguin N, Gupta P, Zhang Z, Peltonen L. Fabricated citations: an audit across 2.5 million biomedical papers. Lancet. 2026;407:1779-81.
[2] Oza A. Fraudulent citations, blamed on AI hallucinations, are becoming more common in research papers. STAT News [Internet]. 2026 May 7 [cited 2026 May 20]. Available from: https://www.statnews.com/2026/05/07/lancet-study-finds-steep-rise-fraudulent-citations-academic-papers/
[3] Naddaf M, Quill E. Hallucinated citations are polluting the scientific literature. What can be done? Nature. 2026;652:26-9.
[4] Retraction Watch. One in 277 PubMed-indexed papers in 2026 shows fabricated references, says analysis [Internet]. 2026 May 7 [cited 2026 May 20]. Available from: https://retractionwatch.com/2026/05/07/one-in-277-pubmed-indexed-papers-in-2026-shows-fabricated-references-says-analysis/