arXiv began enforcing a strict new policy this week: submit a paper containing AI-hallucinated citations and face a one-year ban from the platform. After the ban expires, all future submissions must clear peer review at a reputable venue before posting. The policy responds to a tenfold increase in fabricated references since 2023.
Hallucinated Citations Increased Tenfold in Three Years
The scope of the problem has grown dramatically:
- 2023: Roughly 1 in 2,828 papers contained a hallucinated reference
- 2025: Climbed to 1 in 458 papers
- Early 2026: Reached 1 in 277 papers
- This represents a tenfold increase in just three years
Major Conferences Discovered Over 100 Fake Citations
The problem has infiltrated top-tier academic conferences:
- NeurIPS 2025: Over 100 hallucinated citations discovered spanning 51 accepted papers
- ICLR 2026: 20% of sampled papers (60 out of 300 papers from 20,000 total submissions) contained at least one AI hallucination
Nature published an article titled "Hallucinated citations are polluting the scientific literature. What can be done?" highlighting the broader academic concern.
Harsh Penalties Reflect Severity of Academic Integrity Threat
The policy represents a severe penalty for what might be accidental reliance on AI writing tools. A one-year ban plus permanent peer-review requirements afterward marks a major policy shift for the scientific community's most important preprint server.
Computer science professor Thomas Dietterich shared the story on Twitter, and it reached 240+ points on Hacker News with 79 comments on May 14, 2026. The policy highlights the growing tension between AI assistance in academic writing and scholarly integrity.
Key Takeaways
- arXiv now enforces a one-year ban for papers containing AI-hallucinated citations, with permanent peer-review requirements afterward
- Hallucinated references increased tenfold from 1 in 2,828 papers in 2023 to 1 in 277 papers in early 2026
- NeurIPS 2025 discovered over 100 hallucinated citations across 51 accepted papers
- ICLR 2026 found 20% of sampled papers contained at least one AI hallucination
- The policy represents a major shift for scientific publishing's most important preprint server