arxiv.org web signal

Audit of 111M references finds 146,932 fake 2025 citations

By Published Updated

TL;DR

  • Researchers audited 111 million references across 2.5 million papers on arXiv, bioRxiv, SSRN, and PubMed Central to find non-existent citations at scale.
  • They estimate 146,932 hallucinated citations in 2025 alone, concentrated in fields with rapid AI uptake and in small or early-career author teams.
  • Fake citations disproportionately credit already prominent and male scholars, and existing preprint moderation catches only a fraction, the paper says.

A new preprint on arXiv estimates 146,932 hallucinated citations in scientific papers in 2025 alone.

The authors audited 111 million references across 2.5 million papers on arXiv, bioRxiv, SSRN, and PubMed Central. Scientific citations, they write, are "a uniquely verifiable object" for testing the real-world magnitude of the hallucination problem.

"We find a sharp rise in non-existent references following widespread LLM adoption," the paper reports. The errors cluster in fields with rapid AI uptake, in manuscripts carrying "linguistic signatures of AI-assisted writing," and among small and early-career author teams.

The credit those fake references distribute runs the other way. Hallucinated citations "disproportionately assign credit to already prominent and male scholars," the paper says, "suggesting that LLM-generated errors may reinforce existing inequities in scientific recognition."

Preprint moderation and journal publication processes, the authors add, "capture only a fraction of these errors." Five of the AI researchers we track had already shared the paper's link by the time we picked it up.

Shared on Bluesky by 5 AI experts

Originally reported by arxiv.org

Read the original article →

Original headline: [2605.07723] LLM hallucinations in the wild: Large-scale evidence from non-existent citations