Generated by Codex with GPT 5.6 Sol High

Techmeme surfaced arXiv’s October 1 announcement, “Fair Moderation, Equitable Access, and AI: arXiv’s Updated Rate Limit Policy”. The research repository is imposing a universal ceiling of two submissions per person per calendar month, plus no more than three active submissions at once. What looks like a narrow moderation rule is also an early institutional response to a larger shift: AI has made producing plausible research prose much cheaper, while evaluating whether that work is worthwhile remains stubbornly human.

The workload change is stark. arXiv received 40,363 submissions in September 2026, nearly twice the 20,569 submitted in September 2024 and more than four times the 9,869 submitted in September 2016. Those papers generated almost 9,000 support tickets. The computer-science artificial-intelligence category grew more than sixfold in two years, while submissions in other categories roughly doubled.

arXiv does not reject AI-assisted research as a category. Its policy permits authors to use AI if they disclose it and the resulting work is original, significant, self-contained, and relevant. The problem is an increase in papers that moderators say fail those standards: thin results, narrow fragments of a larger project, and dense AI-written manuscripts. A small group of high-volume submitters can therefore consume a disproportionate share of volunteer attention and delay other papers for days or weeks.

The new rule prices that scarce attention in submissions rather than publications. A rejected paper still uses one of the author’s two monthly slots because a moderator had to assess it. A paper deleted before announcement does not count. The cap follows the person who uploads the manuscript, not every co-author, so collaborators must coordinate who submits. It applies across subject areas, which prevents an author from bypassing the limit by spreading papers among categories.

This is a rate limit, not an AI detector. That distinction matters. Automated screening would have to infer authorship or quality from text and could misclassify careful work, especially as generated prose becomes harder to distinguish from edited prose. A simple quota is legible and immediately reduces worst-case load without requiring arXiv to decide whether a particular paragraph came from a model. It also encourages authors to bundle related findings and choose their strongest work before asking volunteers to review it.

The policy’s evidence has limits. The overall submission counts are direct, but arXiv does not quantify what share of the increase is AI-generated, low quality, or caused by conventional academic incentives such as “publish or perish.” The two-paper ceiling also applies to prolific researchers whose work is legitimate, while the submitter-only design leaves room for large collaborations to rotate uploaders. As a stopgap, it shifts costs back toward authors; it does not fix the incentives to split work, increase moderation capacity, or establish whether submission quality improves.

The deeper lesson reaches beyond scholarly publishing. Generative systems can multiply artifacts faster than trusted institutions can evaluate them. When production becomes cheap but judgment does not, open systems need admission controls that protect human attention. arXiv’s experiment makes that bottleneck explicit: the scarce resource is no longer the ability to write a paper-shaped object, but the expert time required to decide whether it advances knowledge.