Tens of thousands of scientific papers published in 2025 contain citations to research that doesn't exist. They're hallucinations, generated by AI writing assistants that researchers increasingly rely on.
A new analysis published in Nature reveals the scale of the problem: AI tools like ChatGPT and other large language models have been confidently citing papers that were never written, by authors who never existed, in journals that may be real but never published the claimed work.
Here's what makes this particularly insidious: the fake citations look completely plausible. They follow proper formatting conventions. They include realistic author names, journal titles, volume numbers, page ranges. Some even have DOIs that lead nowhere, or worse, to entirely different papers.
"This isn't just a footnote problem," says Dr. Guillaume Cabanac, a computer scientist at the University of Toulouse who has been tracking these hallucinated references. "When citations are wrong, the entire edifice of scientific reproducibility begins to crumble."
The problem emerges from how large language models work. These systems are prediction engines, trained on vast corpuses of text. When asked to provide a citation, they don't search a database of real papers—they generate text that statistically resembles a citation based on patterns they've learned.
Sometimes they get lucky and hallucinate something that happens to be real. Often they don't.
Researchers have identified several warning signs of AI-generated citations:
Suspiciously perfect formatting - Every citation follows the exact same style, with no variations in punctuation or abbreviation that naturally occur when humans compile references from different sources.
Clusters of fake citations - Multiple non-existent papers appearing together in reference lists, often with similar themes or methodologies.
Broken DOIs - Digital object identifiers that don't resolve, or point to completely unrelated papers.
