By ByteBot
arXiv this week began enforcing a blunt new policy: submit a paper containing AI-hallucinated citations — references to papers that don’t actually exist — and you’re banned from the platform for a year. After that year, all future arXiv submissions must first clear peer review at a reputable venue before you can post. The announcement is trending on Hacker News today as researchers and developers digest what it means. It arrives as data shows hallucinated citations have risen tenfold since 2023, reaching 1 in every 277 papers in early 2026.
The Numbers Behind the arXiv Ban
The rate of papers with fabricated citations has exploded in near-perfect correlation with AI writing tool adoption. In 2023, roughly 1 in 2,828 papers contained a hallucinated reference. By 2025 that had climbed to 1 in 458. By early 2026, it was 1 in 277 — a tenfold increase in three years. A Lancet study published this month analyzed over 2 million papers and 97 million citations, finding roughly 4,000 fabricated citations across 2,800 papers in the sample alone.
The NeurIPS 2025 incident put a face on the numbers. GPTZero scanned 4,841 accepted NeurIPS papers and found 100+ hallucinated citations across 53 papers — papers that each beat out 15,000+ other submissions and cleared review by at least three human reviewers. Reviewers, it turns out, almost never check whether cited papers actually exist. The problem is systemic, not anecdotal.
Related: METR AI Survey: Developers Claim 2x Gains, Data Disagrees
What the arXiv Policy Actually Does
The penalty is two-pronged. First, a 1-year ban from submitting to arXiv entirely. Second — and this is the part with real teeth — after the ban ends, the author must have all future submissions accepted at a reputable peer-reviewed venue before posting to arXiv. For researchers in CS and AI who rely on arXiv for rapid preprint dissemination, losing that preprint-first access is a significant career disruption. Most papers go to arXiv the same week they’re submitted to conferences; having to wait for peer review can mean being scooped by months.
arXiv’s stated rationale leaves no wiggle room: “By signing your name as an author of a paper, each author takes full responsibility for all its contents, irrespective of how the contents were generated.” The policy positions hallucinated citations not as an AI failure, but as an author failure. That framing is deliberate.
Two Camps, One Debate
The Hacker News discussion has split predictably. The pro-ban position is direct: if you cannot verify that the references in your own paper actually exist, you are not ready to publish research. Northwestern University professor Mohammad Hosseini put it plainly: “Citation practices are changing…people simply use their hunches to prompt ChatGPT, and that is not a healthy practice.” Supporters argue that citation verification is table-stakes scholarship, not an unreasonable burden.
The anti-ban camp focuses on proportionality and intent. Critics note that fraud traditionally requires intent to deceive — hallucinating one citation out of eighty while using an AI writing tool is closer to negligence than deliberate misconduct. The permanent post-ban peer-review requirement, they argue, is disproportionate for a first offense. However, this argument faces a structural problem: the whole point of using AI writing tools is that the author trusts their output. Choosing not to verify is a choice, not an accident. “My calculator gave the wrong answer” is only a defense if you don’t know calculators can be wrong — and by 2026, everyone knows LLMs hallucinate citations.
What Developers Should Take Away
The arXiv policy applies to arXiv submissions. However, the underlying failure mode — LLMs inventing plausible-sounding references the author doesn’t verify — is not limited to academic papers. Developers writing technical documentation, blog posts, README files, or conference talks face the same risk whenever they ask an AI to “add citations” or “include supporting links.” The same models that fabricate academic citations will fabricate GitHub repo links, documentation URLs, and author names. The NeurIPS irony — the world’s top AI conference accepting papers with AI-hallucinated citations — is a useful reminder that even experts in the field are not immune.
Related: AI Code Quality Crisis 2026: The Hidden Cost of Productivity
Key Takeaways
- arXiv is now banning authors for 1 year if they submit papers with AI-hallucinated citations, followed by a mandatory peer-review gate for all future submissions
- Hallucinated citations in academic papers rose tenfold between 2023 and early 2026, from 1 in 2,828 papers to 1 in 277
- At NeurIPS 2025, 100+ fabricated citations slipped through in 53 papers that cleared human peer review — reviewers don’t check citation lists
- The policy frames this as author responsibility, not AI failure — signing the paper means owning its contents
- The same hallucination risk applies to any LLM-assisted writing that includes external references, not just academic papers
