ArXiv, the preprint server where researchers post scientific papers before formal peer review, is now banning authors who submit work with obvious signs of unchecked AI generation. The policy targets what the platform calls "incontrovertible evidence" of careless AI use - specifically, hallucinated references (citations to papers that don't exist, fabricated by a language model) and LLM meta-comments, the instructional filler text that AI writing tools sometimes leave behind, phrases like "[insert relevant statistic here]" or "[this section needs expansion]."
The bar ArXiv set is deliberately high. Researchers who use AI tools to help write or edit papers are not the target - the ban applies when authors clearly never read what the AI produced before submitting it. Hallucinated citations are the clearest tell: a paper that cites a 2023 Nature study that simply does not exist suggests the author copied AI output without checking a single reference.
This matters beyond academic housekeeping. ArXiv hosts over two million papers and is the primary distribution channel for AI research itself - most papers from Google DeepMind, Anthropic, Meta, and university labs appear there first, often before any journal peer review. If AI-generated slop pollutes that pipeline, it corrupts the research that other researchers, journalists, and product teams build on. Bad citations get cited downstream. Nonexistent findings get treated as established results.
The move also signals a broader reckoning that academic institutions have been slow to reach. Journals have individually caught and retracted papers with obvious AI artifacts - the phrase "Certainly! Here is your text" appearing verbatim in published research became an embarrassing recurring news item throughout 2024 and 2025. ArXiv's approach is pre-emptive: catch it at submission, not after citation damage spreads.
For researchers, the practical message is simple. Using AI to help draft, edit, or summarize is not the issue. Submitting output you haven't read is grounds for losing access to the most-read preprint server in science.