AI-Fabricated Citations In Over 2,800 Biomedical Journal Articles

By 2026, Approximately One in 277 Papers Contained Presumably AI-Fabricated Citations

A team of researchers from Columbia University and the University of Eastern Finland embarked on a study to uncover AI-generated content. They created an automated reference verification system to sift through the vast number of papers—2,471,758 to be exact—with 125,615,773 references published in PubMed Central’s Open Access from January 1, 2023, to February 18, 2026.

They compared the citations within these papers against actual bibliographic records to identify discrepancies. Further AI assistance came from a large language model named Claude 3.5 Haiku from Anthropic, which helped distinguish between genuine errors and deliberate fabrications. References missing from databases such as PubMed, Crossref, OpenAlex, and Google Scholar were marked as fabricated.

In 2023, one in every 2,828 papers had at least one fabricated citation. By 2025, this figure had alarmingly risen to one in 458 papers, and by the first seven weeks of 2026, it surged to one in 277 papers, indicating a more than 12-fold increase in a short span.

A 2025 Paper Had 60 Percent of Its Citations Presumably AI-Fabricated

An open access oncology journal paper published in 2025, discussing ureteroileal anastomotic techniques, had an astonishing 60 percent of its references fabricated. This is just one example among many papers with a high number of fabricated references. In total, 246 papers had three or more fabricated citations.

The research identified certain authors who consistently produced papers with fabricated references. For instance, two authors published 11 papers in a surgical journal in 2025 that collectively contained 15 fabricated references. The study also pointed to “paper mills,” or groups that mass-produce academic papers, with review articles showing the highest fabrication rates—57 percent higher than other types of articles.

AI Large Language Models Pose a High Risk for Hallucinations

Without direct communication with the authors, the reasons behind the fabricated citations remain speculative. However, it is believed that these fabrications are not intentional but rather the result of hallucinations from large language models (LLMs).

The increase in fabricated citations coincides with the growing use of AI platforms like ChatGPT, Claude, and Perplexity, which utilize LLMs. These models are prone to hallucinations as they associate data without critically evaluating its accuracy. I’ve discussed these AI hallucinations in Psychology Today under the article “A Funny Bone to Pick.”

AI-Fabrications Will Continue to Challenge Scientific Publishing

The issue of fabricated citations is unlikely to disappear and may worsen as researchers increasingly depend on AI to draft their papers. The proliferation of profit-driven scientific journals, which expect scientists to review papers for free, exacerbates the quality-control problem. Many established scientists are reluctant to engage with these journals due to the lack of compensation.

One proposed solution is to develop AI tools to detect inaccuracies, especially those generated by AI. However, whether publishers will invest in such tools and their potential accuracy remain uncertain.

Scientific publishing faces a potential reckoning. The view of journals as profit-generating entities has led to a surge in publications charging researchers hefty fees for publication. Simultaneously, funding for scientific research is being reduced, leaving fewer resources for researchers to pay these fees. This situation may drive more researchers to cut corners and rely on AI, even before these tools are fully vetted for accuracy. As the title of a popular 2003 rom-com suggests, “something’s gotta give.”