arXiv — the preprint server that AI research runs on — capped author submissions at two per calendar month this week, after September 2026 set a record 40,363 submissions. That figure is double September 2024’s total and quadruple 2016’s. The cs.AI category alone grew sixfold in two years. Moreover, the platform’s volunteer moderation infrastructure, designed for a few thousand papers a month, generated roughly 9,000 support tickets last month from a system simply never built for this scale. arXiv’s own description of the updated rate limit policy is telling: a “stopgap while we determine what may be the new best practice.” They know it’s not a fix.
The Numbers Behind the arXiv Submission Cap #
The growth trajectory tells the story more clearly than any policy announcement. In September 2016, arXiv received 9,869 submissions. By September 2024, that number had risen to 20,569. This past September it hit 40,363 — generating roughly 9,000 support tickets for volunteer moderators in a single month. Furthermore, the cs.AI category saw more than sixfold growth in just two years, driven almost entirely by the proliferation of large language models that make drafting research papers dramatically faster.
This isn’t a slow-burn trend that snuck up on anyone. arXiv’s submission volume has doubled roughly every two years since 2022 — the point at which LLMs became widely accessible to researchers. Additionally, the platform became an independent nonprofit on July 1, 2026, absorbing this doubled volume at the exact moment it lost the institutional cushion of Cornell’s operational support. The timing couldn’t be worse.
What’s Actually Flooding the Platform #
arXiv’s moderators identified three distinct failure modes in the surge. First: “thin papers of narrow scope” — work that would have been a conference poster packaged as a standalone submission. Second: “salami slicing,” where a single piece of work is fragmented into multiple papers specifically to inflate publication counts, a well-known gaming strategy in publish-or-perish academic culture now being turbocharged by AI writing tools. Third: dense AI-written papers submitted without meaningful human review — including papers with hallucinated citations and leftover chatbot instructions in the body text.
Thomas Dietterich, chair of arXiv’s editorial advisory council, noted that “some authors filing dozens of papers each” were “delaying release of everyone else’s work.” This is the core problem: a small number of high-volume submitters impose costs on everyone else’s timelines. Worth noting: arXiv had already escalated once in May 2026, implementing a one-year ban for papers with obvious LLM artifacts. Consequently, the submission cap is the next escalation in a pattern that isn’t stabilizing.
Related: Google Kills Bug Bounty: AI Slop Breaks Open-Source Security
The Cap Is a Band-Aid, Not a Fix #
The policy has its supporters. Rutgers mathematician Alex Kontorovich put it simply: “I’m not going to read your 42 papers this month. Pick two, your very best two.” However, Harvard’s Scott Kominers disagreed sharply, arguing arXiv should “economise on moderation — not on scholarship,” as noted in Times Higher Education’s coverage of the divided reaction. In other words, rate limits don’t distinguish between genuinely valuable work and junk — they just slow everything down equally.
Sabine Hossenfelder offered the sharpest prediction: papers are going to get longer. Researchers who would have submitted two shorter papers will merge them into one bloated submission. As a result, the same content appears in a harder-to-read format with no net improvement in signal quality. Meanwhile, Mark Lewko raised a structural concern: rate limits don’t reduce paper production — they delay public availability. This pushes researchers back toward private correspondence to establish priority, eroding the scientific openness that arXiv was built to enable.
The critics are right. The submission cap treats the symptom without touching the disease: an academic evaluation system that rewards publication counts over research quality, combined with AI tools that make gaming that system trivially cheap. As Daniel Lemire observed, this mirrors Google halting new bug bounty submissions for the exact same reason — volume management fails as a strategy when the underlying incentive to produce volume doesn’t disappear. In contrast, fixing the publish-or-perish culture would address the root cause.
The Pattern Is Bigger Than arXiv #
arXiv is not an isolated case. This month, Google halted new bug bounty submissions because AI-generated slop reports overwhelmed human reviewers. Similarly, Reddit shut down its public API and RSS feeds, partly to combat AI scraper abuse. The dynamic is identical in each case: AI tools make bulk content generation cheap enough that bad actors — and incentivized non-bad actors — can flood human-moderated systems faster than those systems can respond. Therefore, rate limits are a defensive crouch, not a structural answer.
There’s a particular irony worth naming. arXiv is where AI researchers publish AI research. The tools being researched inside the platform are the same tools breaking the platform from outside it. For example, a paper currently on arXiv is titled “Stop DDoS Attacking the Research Community with AI-Generated Survey Papers.” Researchers are using the platform to publish warnings about what’s happening to the platform.
Key Takeaways #
- arXiv’s new policy (effective October 1, 2026): 2 submissions per author per month maximum, 3 active submissions at any time; rejected papers count toward the monthly quota
- The trigger: 40,363 submissions in September 2026 — doubled in two years, driven largely by the cs.AI category and AI-enabled bulk paper production
- The cap is explicitly a “stopgap” in arXiv’s own language — not a long-term solution, and critics who say it misses the point are making the stronger argument
- The real problem is structural: publish-or-perish incentives plus AI writing tools create an arms race that rate limits alone cannot stop
- The same pattern — AI content flooding human-moderated systems — has hit Google’s bug bounty program and Reddit’s public API this month