The Archive That Throttled

On October 1, arXiv introduced a limit of two submissions per calendar month across its categories, alongside an existing ceiling of three active submissions. The open research repository describes the change as a stopgap: time to improve moderation tools and procedures while advanced AI changes how much material researchers can produce.

Matthew Gault's report for 404 Media frames the move as a response to AI slop. The archive's own announcement supplies a more useful distinction. AI assistance is permitted under its standards; an unlimited claim on volunteer attention is not. That difference matters both to people producing worthwhile research with new tools and to the people asked to sort it from everything else.

This is not a story about an upload button running out of enthusiasm. It is about the labor behind the button—and a policy that protects that labor by rationing access to it.

Count the work, not just the papers

According to arXiv's announcement, it received 40,363 submissions in September 2026, compared with 20,569 in September 2024 and 9,869 in September 2016. Last month's submissions generated almost 9,000 support tickets for staff and moderators. The institution reports a greater-than-sixfold increase over two years in its cs.AI category, not in computer science as a whole.

These are arXiv's reported internal figures, not an independent audit. They establish the scale of the intake the institution says it faces. They do not establish what proportion of that increase came from AI, how many submissions were poor, or how much moderator time each kind of paper consumed. A submission count is not a quality assessment wearing a spreadsheet.

The announcement does describe what moderators are seeing: thin papers with narrow scope, “salami” submissions that divide one work into several smaller papers, and dense AI-written papers. Thomas Dietterich, chair of arXiv's Editorial Advisory Council, says a relatively small proportion of authors are submitting many low-quality papers and taking a disproportionate share of moderator time. He says this can delay quality papers for days or weeks.

That is an institutional diagnosis worth taking seriously, while keeping its limits visible. The announcement does not provide a measured breakdown of the small group, its submissions, or the delays it causes. Nor does it demonstrate that every additional paper is an unwanted one. arXiv itself acknowledges that advanced AI tools can make parts of research faster and easier.

The policy therefore addresses a real reported capacity problem without resolving the entire argument about AI's scientific value. It does not need to resolve that argument to justify protecting staff and volunteers. But the distinction becomes essential when asking whether this particular restriction is well targeted.

What the ceiling actually measures

The most consequential word in the policy is “submitter.” In its FAQ, arXiv explains that uploading a paper with several co-authors counts against the uploader's allowance, not every co-author's. This is not a rule limiting every researcher to appearing on two papers a month. It limits the submission activity attributed to the person doing the uploading.

The two ceilings also do different jobs. Two submissions per calendar month constrain the arrival of new work. Three active submissions constrain how much unfinished submission activity one submitter can have in the system. arXiv says the three-active limit has existed since 2024; presenting both numbers as newly invented would exaggerate the change.

A paper rejected during moderation still counts toward the monthly allowance. That is deliberate: the resource being rationed is attention spent on submissions, not the number of papers ultimately announced. Someone who submits two rejected papers must wait until the next month before submitting again.

There are further operational details. Papers held over from the previous month do not consume the new month's two-submission allowance, but they still occupy active-submission slots. A paper deleted by its submitter before announcement does not count toward either limit. The moderation policy separately specifies a replacement limit of one per week after version five, distinguishing revisions from the new-submission allowance.

These details make the mechanism legible. Imagine a submitter entering a new month with two papers still on hold. The calendar has renewed their monthly allowance, but only one of the three active slots remains available. The example follows from the rules; it is not evidence that any particular researcher's work has been delayed this way.

That interaction matters because the submitter does not control moderation speed. A constraint intended to protect the queue can also make an individual's available capacity depend on how slowly the queue clears. The archive is allocating a scarce service, not merely asking everyone to write less.

Moderation is not scientific certification

arXiv's content-moderation policy explicitly says moderation is not peer review. Its volunteer subject-matter experts assess whether submissions meet the repository's standards and belong in its categories. Posting a paper does not mean the archive has certified every conclusion, proof or experimental result.

This prevents two convenient errors. The first is treating an arXiv listing as a scientific warranty. The second is treating a stricter intake policy as a substitute for the scrutiny that research still needs after it becomes available. A smaller queue can give moderators breathing room. It cannot, by itself, establish that the papers emerging from that queue are correct.

The distinction also explains why a number alone is a blunt instrument. A monthly cap can stop one submitter from sending a third poor paper. It can equally stop a third substantial contribution. It does not inspect either paper's contents. The hope is that authors will select their strongest work and the remaining moderation workload will become manageable; the announcement states that hope, not a demonstrated outcome.

arXiv's AI rules place responsibility elsewhere too. Authors must report significant use of sophisticated tools, including text-to-text generative AI, consistent with their field's methodological standards. Each author remains responsible for the contents bearing their name, including errors, misleading material and incorrect references generated by a tool. Generative AI tools should not be listed as authors.

That is a stronger standard than a blanket sneer at machine-assisted prose. A useful tool can help someone communicate valid work; a fluent tool can also make unsupported work look finished. Disclosure helps a reader understand the process. It does not repair a false claim. Neither a human byline nor an AI disclosure is a receipt for verification.

Equal numbers, unequal room

A universal rule has an obvious administrative advantage: it is easier to explain than an opaque collection of individual restrictions. arXiv says rate limiting was already established policy, previously left primarily to moderator discretion. The new ceiling makes part of that rationing explicit.

But equal allowances for submitters do not necessarily produce equal room for research groups. The FAQ expressly asks co-authors to coordinate their submissions. A collaboration with several eligible people able to submit can distribute that task in a way a solo author cannot. That is a structural implication of the policy, not evidence of abuse or a measured disadvantage to independent researchers.

The same distinction applies to timing. Work does not necessarily become ready in two-paper monthly installments. A researcher may have a backlog, or a conference may create a deadline. arXiv acknowledges both situations and asks authors to plan their submissions around the limits. Planning is a useful instruction when timing is flexible. It is not proof that the underlying constraint disappears when timing is not.

None of this establishes that the cap does more harm than good. Delays caused by overloaded moderators also restrict access, just less visibly. Leaving the upload channel unrestricted can allow the heaviest users to consume attention that others need. “No limit” is an allocation policy too; it simply declines to announce who pays.

The harder question is whether the new rule distributes that cost more fairly. Answering it requires evidence from the people waiting, not just a falling submission total. The archive's announcement promises to monitor effects on authors, readers and the corpus and modify the policy as needed. That promise is the beginning of an evaluation, not the evaluation itself.

The commons includes its maintainers

An open archive is not sustained by permission to read alone. Its usefulness also depends on classification, standards, support and the accumulated work that makes a shared record navigable. If those functions are overwhelmed, nominal openness can coexist with practical exclusion: the paper is theoretically welcome, but nobody has time to handle it.

This is where the coherenceism lens changes the question. Protecting a commons does not mean maximizing every individual's immediate freedom to consume it. It means preserving a shared resource while including the feedback of the people affected by its rules. Moderators belong in that circle. So do careful authors whose papers wait behind noise, and legitimate contributors whose work collides with the new ceiling.

The reciprocity problem is concrete. When a tool reduces the effort needed to produce a submission without reducing the effort needed to assess it, some of the apparent productivity gain becomes work for somebody else. That does not make the tool illegitimate. It does make “more output” an incomplete measure of progress.

Institutions encouraging faster production should therefore ask what they contribute to the infrastructure receiving it. Support for moderation, better-prepared submissions and tools that reduce avoidable checking are possible forms of return. Those are obligations and design directions worth considering—not benefits this policy has already delivered.

Nor should protecting the archive become an excuse to silence criticism of its administration. arXiv's moderation policy allows appeals of classification and rejection decisions, with qualifications, and says appeal decisions are final and provide no feedback. Such a process is not the same as assessing the fairness of a system-wide quota. A person can accept that their paper needs moderation and still reasonably question how access to moderation is distributed.

What would count as working?

A credible assessment would track more than fewer submissions. It would examine turnaround times, support workload, and how often compliant work is delayed by either ceiling. It would ask whether moderator effort is actually less concentrated on problematic submitters, and whether researchers working alone face different practical constraints from larger collaborations.

These are proposed tests, not findings. The sources reviewed here do not establish that the October policy has improved quality, shortened waits or redistributed access fairly. It is too easy to turn the institution's stated purpose into a reported result. “Equitable access” is what the policy must demonstrate, not something its title can install.

There is room for genuine credit. arXiv has named moderator labor as a finite resource, explained several awkward edge cases and called the change a stopgap rather than a final answer. It has also preserved a distinction between permissible AI assistance and work that fails scholarly standards. Those are better starting points than pretending either that all generated text is discovery or that every assisted researcher is a spammer.

The unresolved obligation is to keep the restriction answerable to what happens next. A commons can need a limit without that limit becoming sacred. Preserving the record and preserving participation are connected tasks; neither is completed by making the other disappear from the dashboard.

AI can accelerate the production of things that look ready to read. Until the work of judging them catches up, the promised acceleration comes with a familiar accessory: a waiting list.