A new study has identified nearly 147,000 fabricated citations in scientific papers stored in major research databases, highlighting growing concerns over the use of artificial intelligence tools in academic work.
Researchers linked to Cornell and UCLA examined 111 million references across 2.5 million papers hosted on arXiv, bioRxiv, SSRN and PubMed Central. They identified 146,900 citations that could not be matched to any real publication, many of which appear to stem from AI hallucinations rather than simple errors.
The analysis compared citation patterns before and after 2023, when large language models like ChatGPT became widely available. The team reported a sharp rise in nonexistent references following the widespread adoption of these tools.
"We find a sharp rise in non-existent references following widespread LLM adoption," the authors wrote in their paper, which is itself hosted on arXiv.
Scientific repositories play a central role in research by allowing authors to share papers quickly before formal journal publication. The spread of unverified AI-generated citations across many papers rather than a small number suggests the issue affects numerous researchers who rely on chatbots without thorough checks.
Usha Haley, a professor of management at Wichita State University, described the development as a serious warning in comments to CNET. "Fake or AI-generated citations undermine trust in the scholarly record that provides the foundation on which peer review and cumulative knowledge rest," Haley said. "Disturbingly, this skepticism is now coming from within academia itself and from early career scholars."
arXiv has responded by announcing it will ban authors who submit work containing hallucinated citations or other unchecked AI content. Steinn Sigurdsson, the organization's scientific director, warned of broader damage to the scientific record.
"The corpus of science is getting diluted. A lot of the AI stuff is either actively wrong or it's meaningless. It's just noise," Sigurdsson told CNET. "It makes it harder to find what's really happening, and it can misdirect people."
While researchers have fabricated citations for years, the current scale and ease of generation through AI represent a shift. The study authors noted that some unmatched citations were likely typos, but the post-2023 increase points strongly to automated tools producing plausible but false references.
Scientific papers influence fields from technology to medicine, making the integrity of their citations essential. When false references go undetected, they can propagate errors through subsequent research and erode confidence in published findings.
Officials at the affected repositories have begun implementing stricter review processes. arXiv's move follows earlier statements from Sigurdsson about the challenges posed by AI-generated content flooding the system.
Experts continue to stress the need for manual verification when using AI for literature reviews or reference lists. The problem appears widespread rather than isolated to a few bad actors, according to the study's distribution analysis.
Further monitoring of citation patterns in coming months will help determine whether new policies reduce the number of fabricated references entering the scholarly record.