Limitation or failure · General science · Computer science · Social science

Large-scale audit finds hallucinated citations entering scientific literature

An audit of non-existent references across papers in arXiv, bioRxiv, SSRN, and PubMed Central.

Summary

Zhao and coauthors audit scientific references at large scale and report a sharp rise in non-existent citations after widespread LLM adoption. The study links the errors to AI-assisted writing signatures and argues that moderation and publication safeguards catch only part of the problem.

AI role

Large language models are identified as a plausible source of fabricated citations in AI-assisted scholarly writing workflows.

Narrative role

This is a high-signal caution event because it measures a concrete failure mode entering the scholarly record at scale, not just a lab benchmark of hallucination.

Caveat

The work is a preprint and infers LLM involvement from patterns rather than observing every manuscript production workflow directly.