arXiv caps submissions at two per author per month to stem the flood of "AI slop" in preprints

arXiv, the open-access repository where researchers from around the world post preprints before peer review, has announced a new submission-limit policy: from now on, each author may submit only two papers per calendar month and keep a maximum of three active submissions at any one time. The measure, explained in a blog post by the platform itself, responds to a very specific reason: the system is being flooded with content massively generated or assisted by artificial intelligence, and the team of volunteer moderators who review each submission can no longer keep up.

The figures arXiv provides illustrate the scale of the problem. In September 2016 the platform received 9,869 submissions; by September 2024 that figure had risen to 20,569; and in September 2026 it reached 40,363. In other words, submissions have doubled in just two years, and in the computer science category —the one hardest hit by the generative AI wave— they have increased sixfold. For an infrastructure that depends on volunteer moderators rather than large-scale paid staff, that exponential growth is unsustainable without structural changes.

Thomas Dietterich, professor emeritus at Oregon State University and chair of arXiv's editorial advisory board, sums it up in the statement quoted in the article: "The heart of arXiv's process for detecting and rejecting low-quality papers is our many volunteer moderators [...] However, a relatively small proportion of authors are submitting large numbers of low-quality papers and consuming a disproportionate fraction of moderator time. This is unfair to the authors who continue to submit quality papers, whose work can be delayed by days or weeks as a result". The message makes clear that the problem is not the number of researchers using AI legitimately, but a subset that saturates the system and indirectly harms everyone else.

The text distinguishes several ways in which AI is fueling this saturation. On the one hand, many researchers use language models as a legitimate auxiliary tool: to process large data sets, write code or design parts of experiments. On the other, a growing number of authors use AI directly to write the papers themselves. arXiv clarifies that its policy does allow the use of AI as a support tool, provided it is disclosed and the work meets the standards of scholarly interest and of genuine contribution to the field. The problem, according to the platform itself, is that "many of the submissions arXiv is now receiving do not meet these requirements".

Among the symptoms arXiv describes are the rise in "narrow, insubstantial papers", so-called "salami papers" (the practice of slicing a single study into multiple publications to inflate the paper count) and, notably, a marked increase in dense AI-written papers that make it easy to flood the repository with low-value content. The combination of these practices is what has driven up the volume of submissions without any proportional increase in real scientific contributions.

This decision does not come out of nowhere: it is part of a battle arXiv has been waging for some time against spam and low-quality AI-generated content. The article recalls that about a year ago the platform had already stopped accepting "review articles" and "position papers" in the computer science category, expressly citing low-effort papers written with language models as the reason. In January of this year arXiv began requiring new authors to be endorsed by an author already in the system, tightening the previous requirement, which merely verified an academic email address. And in May, the platform went a step further by announcing that it will ban for one year any researchers who submit what it calls "AI slop".

The wording of the announcement itself clarifies that the option to limit submissions was already in arXiv's policy, but until now was left to the discretion of each individual moderator. What changes now is that it is being formalized and applied uniformly to all authors, with the explicit aim of "fairly distributing the incredibly valuable work" of the volunteer moderators.

The arXiv case matters beyond the repository itself because it functions as critical infrastructure for scientific communication, especially in physics, mathematics and computer science, where posting a preprint before formal peer review is common practice and is closely followed by journalists, other researchers and even by AI systems themselves, which are trained on or documented with those texts. If the quality filter degrades under the saturation of automatically generated content, the damage is not limited to the moderators' experience: it affects the overall reliability of what is considered preliminary scientific literature, and it may end up penalizing researchers who submit genuine work but see their submissions delayed because of noise generated by third parties. It is a fairly clear example of how the ease of producing plausible text at scale with generative AI is forcing academic institutions to redesign, with increasingly restrictive measures, access controls that for decades rested on good faith and on the natural scarcity of human writing effort.