Two a month, after forty thousand in September

Since Thursday, anyone who submits a paper to arXiv may submit two in a calendar month and have no more than three under consideration at once. The preprint server announced the limit the same day and called it a stopgap while it works out "what may be the new best practice for authors employing more and more advanced AI tools", and while it improves its moderation tools. The limit counts submissions, not published papers, so a rejected paper uses up one of the two. It applies across every subject category. It falls only on the person who presses submit, not on that person's co-authors, and a paper withdrawn before it is announced does not count.

arXiv's explanation is a set of numbers. In September 2016 it received 9,869 submissions. In September 2024 it received 20,569. This September it received 40,363, "which in turn generated almost 9,000 support tickets for arXiv staff and moderators". Submissions have doubled in two years, the archive says, and in its cs.AI category they have risen more than sixfold over the same period.

The archive's public monthly statistics show how sudden September was. August brought 31,173 submissions, so September added 9,190 in a single month, a rise of 29 per cent. This desk read the series back to its start in 1991: that is the largest month-on-month increase arXiv has recorded. The next largest was 5,755, in March of this year. September was also 51 per cent above September 2025. From January to September arXiv received 270,685 submissions, against 284,486 in the whole of 2025. The remaining gap, 13,801, is less than half of an ordinary month.

What the extra papers are made of is a judgement from arXiv's moderators, and its post sets it out. They are seeing "thin papers of narrow scope", "salami" papers in which one piece of work is cut into several submissions, and "a marked increase in dense, AI-written papers". arXiv's rules allow authors to use AI in support of their research, provided they disclose it and the work meets the archive's standards. Many of the submissions now arriving, the post says, do not. "AI tools are making it easy for authors to flood arXiv and other repositories with these low-value papers."

Every submission passes a volunteer moderator before it is posted, and that is the resource being rationed. Thomas Dietterich, an emeritus professor at Oregon State University who chairs arXiv's editorial advisory council, put the problem in terms of fairness: "a relatively small proportion of authors are submitting a large number of low-quality papers and consuming a disproportionate fraction of the moderators' time." He added that "this is unfair to authors who continue to submit quality papers — their papers can be delayed for days or weeks as a result."

Before the change, moderators held back submitters they judged to be well above a "practical limit" on how fast independent work could be produced. arXiv's post says that, with AI tools becoming more advanced and accessible, the limit is "now up for debate". A fixed number replaces the judgement. It is blunt by design: it measures how much a person submits because, at 40,000 papers a month, the archive cannot measure quality at the door. A research group with many members can still post plenty, since each submitter has an allowance. One prolific author with a capable model cannot.

The archive is facing this as it becomes an independent nonprofit. On 23 September it announced $17.2 million in commitments over three to five years from Simons Foundation International, XTX Markets and the Siegel Family Endowment. Part of that money is earmarked for technical work "related to the management of AI-generated content". arXiv says it will watch how the limit affects submission rates and moderation load, and change it as needed. The first full month under the rule is October. The archive publishes its count when the month ends.