References in scholarly articles serve to underpin the claims and rationales that research builds upon, forming the foundational network for scientific and medical advancement. While this citation system drives progress, it also presents vulnerabilities, ranging from citation manipulation to the emergence of paper mills. A newer and increasingly alarming issue is the rise of fabricated references.
In a study presented in May 2026, Maxim Topaz of Columbia University and his collaborators demonstrated that fabricated references in biomedical literature jumped from roughly four per 10,000 papers in 2023 to 51.3 per 10,000 by the end of 2025.1 The researchers subsequently expanded their investigation to physics, computer science, materials science, engineering, and chemistry. Although such papers represent a small fraction of total citations, the rate of appearance has risen steadily across these fields from 2023 to 2026.2 The findings, reported in the Proceedings of the National Academy of Sciences, underscore the need for publishers and the scientific community to detect and eliminate these spurious sources.
Fabricated References Are Increasing Across Disciplines
Topaz’s team gathered a sample of 500,000 studies across the five fields, indexed in OpenAlex between January 2023 and 31 March 2026. After extracting citation data, they compared reference titles with the identifiers provided via DOIs. A large language model helped filter out false positives caused by formatting variances. When a DOI link did not match the cited title, the researchers searched for the title across seven scholarly databases and search engines. Entries that yielded no matches were flagged as likely fabricated.
Across 21.6 million references, the team identified 1,135 suspect citations originating from 619 papers in the five disciplines. Most of these articles contained one or two fabricated references, but 27 studies listed five or more. While the absolute number remains modest, Yian Yin of Cornell University, who studies AI’s impact on science, warns that even a small proportion can reverberate through the scientific ecosystem. “Even if it starts with a very small fraction, through these mechanisms you might see an even longer impact on the entire scientific ecosystem,” he observed.
The average prevalence of published articles with at least one fabricated reference in the citations had risen almost four‑fold since 2023 in the first quarter of 2026.
Erin Lemieux/Adapted from Topaz et al, 2026, CC‑BY‑NC‑ND
The incidence of fabricated references varied by field. Physics showed the highest rate—25.6 affected papers per 10,000 studies—while chemistry recorded the lowest at 4.3 per 10,000. The authors suggest that physics articles often incorporate preprints, which may escape scrutiny during editorial review.
Overall, the proportion of papers with fabricated references climbed from 6.3 per 10,000 in 2023 to 23 per 10,000 in Q1 2026. “The trend is concerning to me,” Topaz remarked. He added, “This is the tip of the iceberg,” noting that the current analysis focused on citations with verifiable DOIs. The team is now investigating references lacking DOI identifiers, which they suspect harbor a much larger share of fabricated content.
Reliance on AI Is Likely Driving the Increase in Fabricated References
Although the study did not directly attribute the rise to AI, the temporal alignment suggests a possible link. Topaz’s group recently posted a preprint on arXiv showing that many large language models still generate spurious citations. Yin, commenting on the work, said it complements his own preprint demonstrating a surge in hallucinated references following the introduction of advanced language models. He cautioned that assessing fabricated‑reference rates prior to 2023 would strengthen confidence that AI is the driving factor.
A recent preprint indicating that roughly 89 % of open‑access articles in PubMed Central bear evidence of AI writing further emphasizes the issue. Topaz stressed that authors must recognize that AI tools can err, including in citation generation. “A reference is a promise to the reader that the source exists,” he noted, “and with this growing trend of AI‑assisted papers, that promise is not being kept.”
