arXiv Is Rate Limiting Authors Following a Flood of AI Slop Submissions

What it claims

arXiv now limits each submitter to two submissions per calendar month, across all subject categories. This is on top of the limit of three active submissions at a time that has applied since 2024. arXiv calls the cap a stopgap while it improves moderation and updates its AI-use guidance.

The stated reason is that a small share of authors submitting many low-quality papers takes up a disproportionate share of volunteer moderator time. That delays quality papers by days or weeks. The specific patterns moderators report are:

• thin, narrow-scope papers
• "salami" papers, where one work is split into several smaller papers
• dense, AI-written papers

Evidence

• September brought a record 40,363 submissions. The page compares this with 20,569 in September 2024 and 9,869 a decade earlier.
• That month also produced almost 9,000 support tickets for staff and moderators.
• In cs.AI, submissions grew more than sixfold in two years.
• Ralph Wijers, then chair of the editorial council, estimated the rejection rate rose from a historical average of about 4% to roughly 10–12%. He said moderators saw no change after ChatGPT launched in 2022 through 2024, and that the surge began in early 2025, first in CS and then in other fields.
• Wijers said "almost all the AI generated rubbish comes from first-time submitters."

How to use it

Rules that affect how you plan submissions:

• The quota belongs to whoever submits. Co-authors use none of their own allowance when someone else submits. The FAQ explicitly encourages co-authors to coordinate who submits what. A team with several qualified submitters can spread papers across them; a group with one or two submitters cannot.
• Rejections count toward the monthly two.
• AI tools are allowed if their use is disclosed and the work meets arXiv's standards.
• CS review articles and position papers are no longer considered unless they have passed peer review and been accepted at a journal or conference. Before this, moderators could admit them at their discretion.
• New submitters need endorsement. An institutional email alone no longer qualifies. You need either:
• an institutional email plus prior authorship on an accepted paper in the relevant endorsement domain, or
• a personal endorsement from an established arXiv author in that domain.

This started in mathematics and was extended to all sections in January.

• Unchecked AI output is penalised. The penalty is a one-year submission ban for all listed authors. After the ban, their future submissions must pass peer review first. The examples given are hallucinated references and leftover chatbot instructions. arXiv leadership confirmed to Ars Technica that this is official policy.
• Plan backlogs and conference deadlines around the cap. Once you hit the limit, you wait until the next calendar month.

Limits

• Revisions are unresolved. The FAQ does not say whether revisions to existing papers use up monthly slots. This matters for conference proceedings published through arXiv and for journals that rely on arXiv to host revised manuscripts. Check this before relying on arXiv for revision workflows.
• The cap applies to everyone, including established authors with accepted papers. Commenters objected because arXiv's own explanation blames a small group. Proposed alternatives included:
• targeting accounts with high-burden submission patterns
• averaging the limit over a longer window for authors whose papers become ready in bursts
• using AI-assisted triage, such as detecting overlapping papers and checking references
• Other preprint servers are stricter.
• The OSF-hosted generalist preprint server suspended new submissions in August 2025, and the Center for Open Science later extended that suspension indefinitely.
• PsyArXiv moved to moderating papers before publication in August 2025.

If your fallback is a smaller preprint server, it may not take your paper.

• Year labels are ambiguous. The page gives relative dates ("last October", "this month", "by May") and does not always state the year.