ArXiv Bans Researchers For Unchecked AI Output

Why this is here: A Columbia University study found that the rate of papers with at least one fake reference climbed from one in 2,828 in 2023 to one in 277 in early 2026.

ArXiv, the open-access repository for preprints, will ban researchers for one year if they submit papers with obvious signs of AI-generated content they haven’t verified. Thomas Dietterich, chair of arXiv’s computer science section, announced the policy on Thursday. The penalty applies to submissions containing hallucinated references, chatbot instructions, or fabricated data tables.

The rule doesn’t prohibit AI use for drafting or editing. It targets instances where authors paste language model output without review.

A 2026 study by Columbia University found fabricated citations have risen twelvefold since 2023, now appearing in roughly one of every 277 papers. ArXiv receives thousands of submissions monthly, straining its volunteer moderation system.

The platform is also becoming an independent nonprofit. This change should allow greater autonomy over moderation and fundraising for quality control.

While the rule addresses careless submissions, it won’t catch more subtle issues like fabricated data or scientifically unsound arguments, which still require peer review. The platform continues to refine its approach to maintaining research integrity.