This morning, the same story flashed twice in my feed from different angles: starting October 1, arXiv allows a single author no more than two uploads per calendar month and no more than three papers with “active” status at any one time. The line looks like a technical tweak to one website’s settings. In fact, it is the first clear admission that scientific infrastructure has run out of a resource that, until now, was assumed to be infinite.
arXiv is an archive of scientific preprints—that is, papers posted by authors before peer review. It was built in 1991 by physicist Paul Ginsparg so colleagues would not have to mail one another printouts; now it is the main place where physicists, mathematicians, and computer scientists show their work before it reaches a journal. Reading and uploading are free—that is the entire business plan, and it also contains the reason for the current limit.
Because only the reading and uploading are free. Between draft and publication stands a person—a volunteer moderator, who gives each submission a quick look and decides what to do with it. Their time is not counted in any public spreadsheet.
The numbers arXiv uses to explain its decision rise in a neat staircase. September 2016—9,869 submissions. September 2024—20,569. September 2026—40,363 submissions, a record in the archive’s entire history, twice as many as just two years earlier. Add nearly 9,000 support requests in that same month—that is, submissions on which moderators and staff had to decide something manually.
The cs.AI section—artificial intelligence within computer science—grew more than sixfold over those same two years. The other sections doubled. For scale: in 2019, arXiv accepted around ten thousand papers a month, or 600–800 a day, and that was considered a large volume. Even then, roughly 15% of submissions each day were flagged for additional review, and a decision had to be made before 14:00 U.S. Eastern time, because that was when the next mailing-list issue went out.
The cause of the surge, as arXiv states outright, is generative AI. It is permitted as a tool, provided the author discloses it and takes responsibility for the content. The problem lies elsewhere: the tool has removed the physical limit on an author’s output. Moderators describe three kinds of junk—“thin” papers on a narrow topic, “salami” papers (one study sliced into several little ones), and dense texts written entirely by a model.
The word “moderation” should not be confused here with peer review: a moderator does not check whether the physics is right. They check whether a submission resembles a research paper and whether it has been assigned to the proper category. arXiv does no peer review at all, and its official guidance stresses this in a separate paragraph.
This is how it is organized: a small staff team uses software to check technical requirements, classification, and metadata, while the substantive part is examined by more than 170 volunteers—recognized specialists in their respective subfields. The benchmark from the official description: “a typical day might involve looking at ten papers and take about ten minutes.” A 2019 blog post describes the routine in more detail—less than thirty minutes a day spent sorting through flagged submissions, ideally alongside a colleague from the same section, because in some narrow categories there is only one moderator at all.
In return, volunteers receive: access to fresh work before everyone else, “service to colleagues,” a line in their professional report, and a sense of belonging to the community, as the website puts it. There is no money. The paid part of moderation—editors, developers, and students—appears in the budget: $4.28 million is allocated for personnel in 2025–2026. The people doing the bulk of the work filtering out low-quality submissions do not appear in the expenses.
arXiv raises money from institutions. Membership is open to universities, libraries, and laboratories; the fee depends on the organization’s “global rank,” calculated from the number of submissions made by its researchers over three years. The top hundred pay $5,000 (or $6,500 under the expanded “All Star” tier), the next four hundred pay $2,500, and the rest pay $1,000. Then there are sponsors and foundations, including the Simons Foundation, NASA, NSF, and Schmidt Sciences, plus Cornell’s contribution—space, administrative services, and other infrastructure.
This is where a discrepancy begins, and I will leave it as it is. The membership page says member fees account for more than half of arXiv’s operating budget. The public 2025–2026 budget lists $1.6 million under “gifts and membership contributions,” against total revenue of $8.4 million and expenses of $8.85 million—meaning the year is planned to run at a $465,000 loss. Either “operating budget” on the membership page means something narrower, or one of those two figures is simply wrong. The published materials do not make clear which.
Next comes the crucial point about measurement. Institutional contributions are calculated based on use, and use means downloads and page views—that is, reading. Writing is not counted at all: no budget line is tied to the number of submissions or to minutes of moderation. For the author of a low-quality paper, that means posting it costs exactly zero, while the costs are spread across the queue and other people’s delays.
And the effect has been measured. A paper published in Science in December 2025 (2.1 million preprints, 28 thousand peer-review reports, 246 million interactions with scientific texts) offers two numbers that strike at the whole structure: researchers who take up language models write more—by 23.7% to 89.3%, depending on the field—and the relationship between writing complexity and the quality of the work has reversed. The texts have become more elaborate. The content weaker.
It would be wrong to say arXiv did not see this coming. Back in 2014, Ginsparg described in Nature an automated screen for fake papers: checking word-frequency distributions catches texts generated by the SCIgen program, which has its own “dialect” in function words like “the,” “of,” and “and.” The method is statistical and surprisingly robust. That same—frequency-based—form of checking is mentioned as still operating in 2026.
The difference is that a language model writes fluently. It is that fluency that breaks an entire system built on the quick glance. Ginsparg calls the old regime “an approximation to peer review by a quick look”: an experienced moderator could see from the abstract and introduction that the author had not read the literature in their field and did not understand why their claim was unlikely. Now, he says, “the abstract can look reasonable... even individual paragraphs,” but “you only realize that something is deeply wrong when you try to follow the logic.” Big claims are made in the introduction and never delivered in the conclusions—the text simply skips steps. And the conclusion that matters more than the influx itself: “Subconsciously, we imagine that the quality of prose is a signal a reviewer uses. With language models, that signal may invert.”
arXiv had tightened its “quick look” before. In May 2026, management confirmed the policy: “irrefutable evidence of AI slop” brings a one-year ban on uploading, after which every subsequent paper must first pass peer review at a reputable publication. Irrefutable evidence means fabricated citations and, especially, traces of the model itself in the submitted file—boilerplate notes such as “here is a 200-word summary; would you like me to change anything?” or “the data in the table are illustrative; replace them with real data from your experiment.” A familiar pattern: first an announcement on social media, then a document on the website. Some researchers asked where, exactly, the policy text was, and were told that there were no new rules—only a clarification of old ones. This is how norms change in scientific infrastructure today: first the post, then the regulation.
The quota was opposed not by defenders of junk, but precisely by people who understand something about this system. Scott Kominers of Harvard Business School, an arXiv author with nearly twenty years’ experience, put the central counterargument this way: cut back on moderation, not science; the policy confuses those two things. In his view, the limit selects not the best work but the work that will pass screening more easily, and those are not the same thing. He also acknowledges that moderators have a hard job, and makes a counterproposal: what is needed is a mechanism for allocating their attention, not a ceiling on submissions.
Alex Kontorovich of Rutgers replied in the opposite spirit: “I like this policy. I’m not going to read your forty-two papers this month. Pick two, the best ones.” Astronomer Catherine Heymans was astonished by the premise itself: who wants to upload that much junk anyway?
Both positions converge on the point where the quota is genuinely crude. The real practice in some fields is not salami slicing but an honest series: a conference deadline, five connected results, and some of the work will break into separate papers anyway. Such an author will have to stretch submission over months, and formally the cure is the same as the cure for abuse: spread it out over time. In other words, the tool does not distinguish among disciplines—it distinguishes only by count.
And one more detail from the official FAQ, one that is scarcely quoted in news coverage. The quota is calculated by submissions, not published papers: if two of your uploads are rejected that month, you cannot submit anything else until the month ends; but withdrawing a submission before the daily announcement comes out does not, conversely, use up your quota. The practical cost: two failed attempts exhaust the month completely. That is exactly what arXiv wanted.
Now for what does not add up in the public numbers. The benchmark of “ten papers in ten minutes” assumes that the more than one hundred and seventy volunteers can process the month’s forty thousand submissions. The arithmetic works only when almost nothing is flagged—and that is exactly what broke. The official explanation is that “a relatively small fraction of authors submit many low-quality works and consume a disproportionate share of moderators’ time.” That sounds plausible, but the publications contain not a single figure showing the size of that fraction. What they do contain are figures on the growth in the total number of submissions—and the quota hits the total number, meaning everyone, including those who have never submitted papers by the dozen.
The irony of the arrangement is that the only things easy to count are the things already being counted: downloads, views, the number of submissions. Human attention is measured by nothing that can be placed in a report, so it has been allocated in the crudest possible way—by count. The limit does not fix the cause. It makes visible what thirty years had concealed behind the word free: a free archive always had a cost price; it was simply expressed in someone else’s evening hours.
Ginsparg also says where this leads, and his line is not cheerful. If fluency of text is no longer a signal of quality, readers will begin relying on other signals: the author’s provenance and the institution’s reputation—“ironically counteracting the democratizing effect of language models on scientific production.” The archive, conceived as a way around editorial gates, returns to judging what is written not by the text but by the signature. The quota does not fix quality—it fixes only the queue. And the question it sidesteps has not gone anywhere: how much attention does one paper cost, and who pays for that attention?