Fabricated References Are Appearing More Often Across Science Papers

References in a published study serve to support the claims and rationale that the present work intends to build upon. Over time, these citations create the backbone on which discoveries and medical practice are made.

While this system is intended to promote scientific advancement, researchers have contended with problems of unscrupulous individuals inflating their own citations during peer review and with the rise of paper mills. However, a new challenge has been increasing in reference sections: fabricated references.

In May of 2026, healthcare AI scientist Maxim Topaz and his team at Columbia University demonstrated the extent of this issue in biomedical research articles, showing that the rate of fabricated references increased from four per 10,000 papers in 2023 to 51.3 per 10,000 papers at the end of 2025.¹ Since then, Topaz said, “The question we’ve heard a lot was: Is this just a medicine problem?”

Now, the team has explored the prevalence of fabricated references in physics, computer science, materials science, engineering, and chemistry. While the overall number of papers with apparently fabricated references was a small proportion of the overall citations, the rate of their appearance across these fields has increased between 2023 and 2026.² The findings, published in the Proceedings of the National Academy of Scienceshighlight the importance for publishers and the research community at large to find and remove these invalid sources.

Fabricated References Are Increasing Across Disciplines

Topaz and his team downloaded a sample of 500,000 studies across these five fields indexed between January of 2023 and March 31, 2026, from the scholarly catalog OpenAlex. After extracting the papers’ references, they compared the titles in the references with those that they identified via the included digital object identifier (DOI) link. They used a large language model to help discriminate false positives that arose in their verification pipeline due to formatting differences. If the source at the DOI link did not match the reference in the paper, the team searched for the titles across seven sources, including scholarly databases and search engines. Those entries that did not return results were considered fabricated references.

In total, the team evaluated 21.6 million references and identified 1,135 of them as likely fabricated. These came from 619 papers across all five fields. The majority of these articles had one or two fabricated references out of their full list of citations, but 27 studies contained five or more apparently fabricated references.

Although this total number represents a small fraction of fabricated references from the study sample, Yian Yina computer scientist at Cornell University who studies AI’s impact on science and was not involved in the study, noted that the literature is an ecosystem that researchers rely on to build their own work. “Even if it starts with a small, very small fraction, through these mechanisms, you might see an even longer impact on the entire scientific ecosystem,” he said.

While every discipline had studies that included fabricated references within them, the rates of this impact were not consistent across fields. Physics was the most affected, with 25.6 papers containing a fabricated reference per 10,000 studies. Chemistry had the lowest incidence at 4.3 affected papers per 10,000.

Topaz and his team said that this variance could be because physics studies are more likely to include preprint articles in their references, so unverifiable references may not raise as much alarm during editing.

The researchers also noted that, across fields, the rate of published studies with fabricated references has increased from 6.3 articles per 10,000 in 2023 to 23 per 10,000 papers in the first quarter of 2026. More so than the total numbers of fabricated references, “The trend is concerning to me,” Topaz said.

He continued, “This is the tip of the iceberg.” Whereas in the present study he and his team focused on articles with verifiable DOIs, the researchers are currently investigating references without these identifiers. “This is, we suspect, where much of the other fabricates and a bigger, much bigger portion are hiding.”

Reliance on AI Is Likely Driving the Increase in Fabricated References

While Topaz and his team did not claim that the trend was caused by new AI models since they did not investigate that, Topaz said that they suspect these models are involved because of the coinciding timelines. In addition, his team recently uploaded a preprint to arXiv showing that many models still fabricate references.³

“[It’s] a very interesting study, and it is also very much complementary to our current preprint work,” Yin said. His team’s study on arXiv also shows an increase in hallucinated references in preprint articles with the introduction of large language models. Yin said that the present study’s approach was carefully done but noted that it would have been good to also assess the rate of fabricated references before 2023 to better confirm the role of large language models in the trend. “Ideally, you should expect something that is flat [which] then increases. That will make us much more confident about this AI effect,” he said.

Considering a recent preprint that suggested that 89 percent of open access articles on the biomedical database PubMed Central have evidence of AI involvement in the writing, Topaz said that it is imperative that authors recognize that large language models can and do make errors, including in citations.⁵

He said, “A reference is a promise to the reader that the source exists, and [with] this growing trend of papers across the sciences, that promise is not being kept.”