arXiv Enforces Two-Submission Monthly Cap to Curb AI-Generated Junk Papers

arXiv, the preprint server where researchers post papers publicly before formal journal peer review, began enforcing a hard ceiling on 1 October 2026: two new submissions per calendar month for every submitter, and no more than three active submissions in the queue at any one time. The month before the change, the repository logged 40,363 submissions, a single-month record and roughly double the 20,569 it received in September 2024. Those submissions generated almost 9,000 support tickets for arXiv staff and its volunteer moderators, according to Kat Boboris, the community engagement manager who posted the policy announcement.

The cap applies to every category on the server — physics and mathematics included, though the surge is loudest in computer science and artificial intelligence. arXiv frames the policy as a stopgap: time to reconsider what a sensible submission rate looks like now that generative AI has erased the old physical limits on how fast one person can produce a manuscript, and time to rebuild the moderation tooling around that new reality.

How the cap actually works

Two is the whole monthly budget, not two per category. A submitter who files one paper to cs.AI and a second to high-energy physics has spent the month. Rejected papers count, because in arXiv's wording it is submissions, not announcements, that consume moderator time. The limit attaches to whoever clicks submit rather than to every co-author, so a five-author paper deducts only from the account that filed it. A paper deleted before it is announced costs nothing, and a paper left on hold from a previous month rolls forward without touching the new month's two — though it still occupies one of the three active slots.

The fine print matters more than usual this time:

• Two submissions per calendar month per submitter, shared across all arXiv categories.
• Three active submissions at any time — submitted, under review, but not yet announced — a rule arXiv has had since 2024.
• Rejections count toward the two; a paper deleted before announcement does not.
• The limit follows the submitter, not the co-authors, which lets a group spread uploads across its members.
• arXiv can ask a particular author to submit even less, and after the fifth version replacements are limited.

One wrinkle is already causing confusion: the announcement's FAQ says limits apply only to submitters, while arXiv's moderation policy page phrases the rule per author. Until that wording is reconciled, the safe assumption is that everyone in a lab should behave as though their own account is being counted.

The numbers behind the stopgap

arXiv's own comparison of Septembers tells the story. September 2016 brought 9,869 submissions; September 2024 brought 20,569; September 2026 brought 40,363. That is a 309 percent rise over the decade, with the growth packed into its final two years. The cs.AI category — computer science, artificial intelligence — grew more than sixfold in that window while other categories roughly doubled. The official announcement attributes the surge to a small set of authors filing many low-value papers and asks researchers to submit work that reflects the full scope of a project rather than multiple incremental pieces.

Steinn Sigurðsson, arXiv's scientific director and an astrophysicist at Pennsylvania State University, told Nature that around 300 volunteer moderators are now checking about 1,000 manuscripts a day, with nearly 4,000 submissions waiting to post. “This step is holding up too many legitimate papers,” he said. The pressure arrived weeks after arXiv announced $17.2 million in multiyear commitments, spanning three to five years, from Simons Foundation International, XTX Markets and the Siegel Family Endowment, which frames its support around equitable digital infrastructure, to back its transition to an independent nonprofit. Part of that money is earmarked for technical work on managing AI-generated content, as Research Information reported. Cash, though, does not mint volunteer moderators overnight.

Photo by Brett Jordan on Unsplash

What the moderators are seeing

arXiv does not run peer review. Before a preprint goes public, a moderator checks that it belongs in its category and clears a basic bar of scholarly interest — and lately, moderators are describing a specific pattern. arXiv names “thin” papers of narrow scope and “salami” papers that slice one result into several smaller submissions, plus a marked increase in dense text that reads as AI-written. Thomas Dietterich, chair of the editorial advisory council and an emeritus professor at Oregon State University, wrote that “a relatively small proportion of authors are submitting a large number of low-quality papers and consuming a disproportionate fraction of the moderators' time,” delaying quality papers by days or weeks. On X he added that the data show some authors filing dozens of papers. The Register catalogued the mechanics, including the point that rejected submissions still spend the month's allowance because they already consumed reviewer attention.

This is not arXiv's first tightening. In October 2025 it stopped accepting computer science reviews and position papers unless they had already cleared peer review at a journal or conference. In January 2026 it raised the bar for new authors, who now need an institutional email plus prior authorship on an arXiv paper in their field, or a personal endorsement from an established arXiv author — a change Science reported after the site had trialed it in the math section during December. In May it suspended some researchers for citation hallucination, the AI habit of inventing references that do not exist. In February, Nature described a human-computer interaction researcher who used an AI tool called Prism to draft a paper about an experiment he never ran — in 54 seconds. That is the throughput problem the moderators are up against.

A blunt instrument, and the objections to it

Victor Galitski, a theoretical physicist at the University of Maryland, points to the obvious hole: co-authors can take turns filing, so a group can move more papers than any single member. SocArXiv, the social sciences repository, has gone the other way, capping each author at one article a week. Its director, sociologist Philip Cohen, says a burst of several submissions is usually “a good tip-off for AI slop,” while allowing that dissertation chapters can legitimately arrive three at a time. Siddhesh Zādéy, an epidemiologist at Columbia University, calls the arXiv cap harmless to genuine productivity, since two preprints a month is already a substantial output. The debate, as Nature reports, splits between people who want the rule tightened and people who doubt a quota fixes anything.

Natalie Khalil, founder of the AI-assisted peer review platform Reviewer3, argues that rejection alone relocates the problem: papers arXiv turns away “will simply go somewhere else and become someone else's problem.” Her counter is capacity — tools that verify code reproducibility, check citations, inspect metadata and review author lists — rather than scarcity. That tracks with the broader industry response to machine-generated manuscripts, from AI paper mill screening across publishers to provenance and watermarking standards for AI-assisted work.

What this means for your lab

A collaborator — call him Dr. M — used to funnel every paper from a five-person group through his own account, because he was the one with endorsement history and an institutional address on file. Under the new rules, that habit breaks exactly when a group hits a conference deadline: one account, two monthly slots, and six finished manuscripts that all need posting in the same week. The base rate says most researchers never feel this. A typical active group posts somewhere between two and eight preprints a year, spread unevenly — far below two a month except in deadline clusters. The exception is the burst, and the burst is precisely when the cap bites.

The fix is not creative scheduling. It is making every member of the group able to submit independently, which means each person needs an institutional email bound to their arXiv account and their past papers claimed. That groundwork has to exist before the deadline month arrives. Four practical steps:

• Check who in the group can actually submit; endorsement status takes time to fix, so audit it now.
• Before any conference deadline, assign each finished paper to a specific submitter so no single account exceeds two in that month.
• Treat a rejection as quota spent — fix whatever got flagged before refiling rather than burning another slot.
• Stop slicing one result into several papers; the cap now prices that habit directly.

One more cost has changed. A rejected or delayed submitter can face stricter scrutiny on later papers, and a paper on hold occupies an active slot for days or even weeks. Sending work in early and clean is worth more than it was a month ago.

Photo by Kelly Sikkema on Unsplash

What happens next

arXiv has called the cap temporary and says it will watch submission rates and moderation load before adjusting. Dietterich has said he hopes the limit can be raised or removed once moderation scales, and arXiv is testing large language model-based detection tools that remain some distance from deployment. Part of the new philanthropic funding is earmarked for exactly that work, which sits alongside the wider question of how journals update their generative AI policies for peer review.

The unresolved variable is the one arXiv itself named: what the right submission rate should be when AI keeps raising how fast one person can produce a manuscript. Until that number is settled, the cap is the institution's honest estimate of what its 300 volunteer moderators can absorb. If you have a paper aimed at a November or December deadline, open your arXiv account today and confirm you can submit — and if you are the lab's designated uploader, hand one of your slots to a co-author now, before the deadline month starts.