Researchers at Columbia University have reported a significant increase in the number of fabricated references in scientific papers, particularly within the biomedical field. A study led by healthcare AI scientist Maxim Topaz found that the rate of fake citations surged from four per 10,000 papers in 2023 to 51.3 per 10,000 by the end of 2025. The findings, which also extend to other scientific disciplines, underscore a growing concern in the academic community regarding the integrity of research literature.
Topaz and his team analyzed a sample of 500,000 studies across various fields, including physics, computer science, materials science, engineering, and chemistry, between January 2023 and March 2026. The researchers extracted references from these papers and compared them with titles identified through digital object identifiers (DOIs) and various academic databases. They detected over 21 million references, identifying 1,135 as likely fabricated across 619 papers. Most of these papers had one or two fabricated references, although some contained five or more.
While the number of articles with fabricated references represents a small fraction of the total analyzed, it raises significant concerns. Yian Yin, a computer scientist at Cornell University and not involved in the study, emphasized that even a small percentage of inaccuracies could lead to broader implications for the scientific literature ecosystem. “You might see an even longer impact on the entire scientific ecosystem,” he stated.
The study published in the Proceedings of the National Academy of Sciences revealed distinct variations across disciplines. Physics had the highest incidence of fabricated references, with 25.6 affected papers per 10,000 studies, while chemistry reported the lowest at 4.3 per 10,000. The researchers suggested that this disparity may stem from physics papers often citing preprint articles, where unverifiable references may not trigger scrutiny during editing.
Overall, the study documented an increase in published studies with fabricated references from 6.3 articles per 10,000 in 2023 to 23 per 10,000 by Q1 2026. Topaz remarked, “This is the tip of the iceberg,” indicating that the current investigation focused only on articles with verifiable DOIs. The team is now looking into references lacking such identifiers, which they suspect may contain a larger number of fabricated entries.
The link between the increase in fabricated references and the rise of AI models is a topic of speculation. Although the researchers did not directly investigate this relationship, Topaz noted a correlation with increasing instances of AI-generated inaccuracies. A recent preprint highlighted that 89 percent of open access articles on PubMed Central showed evidence of AI’s involvement in writing, further complicating the issue of reliability in scientific references.
“A reference is a promise to the reader that the source exists,” Topaz said, underscoring the need for authors and publishers to ensure the integrity of citations as fabricated references become a more prevalent issue in scientific research.
