arXiv’s new submission cap is an invoice for labor everyone pretended was free.
On October 1, the open-access research repository began limiting each submitter to two papers per calendar month and three active submissions at a time. A rejected paper still consumes one of the month’s two slots because the cost enters the system at submission—not publication.
That detail is the story.
In September 2024, arXiv received 20,569 submissions. In September 2026, it received 40,363, generating almost 9,000 support tickets for staff and volunteer moderators. Submissions in its artificial-intelligence category increased more than sixfold over those two years.
arXiv does not prohibit responsible AI assistance. Its policy allows authors to use AI when they disclose that use and the work still meets scholarly standards. The repository says the expensive failure mode is a flood of thin, fragmented, and dense AI-written papers that consume disproportionate moderation time.
The writing looked cheap to the submitter.
The review was not.
The Draft Is Not the Product
Organizations often price AI writing at the moment of generation.
How many minutes did the prompt save? How many pages did the tool produce? How much less did the organization spend on the first draft?
That accounting stops exactly where the work becomes consequential.
Start with who is doing the work. Reviewers check the draft against sources, subject-matter experts test the instructions and trace the claims, and editors surface risks and decide whether the document deserves institutional trust. Future maintainers inherit it when the facts change.
Faster generation does not eliminate those costs—it moves them.
At arXiv, they move to staff and volunteer moderators. Inside an organization, they move to technical writers, editors, subject-matter experts, security reviewers, support teams, and future maintainers.
Calling the draft cheap does not make the information system cheap. That price simply excludes the people doing the rest of the work.
The Queue Shows Whose Work Was Invisible
The review queue makes misunderstood labor visible.
Yesterday I wrote that the best documentation is easy to miss because its value often appears somewhere else. Support sees fewer tickets. Implementation sees less rework. New employees need fewer interruptions.
This story exposes the inverse: when leaders label AI-generated writing cheap, its costs appear somewhere else—in specialist interruptions, support volume, corrections, and maintenance debt.
In both directions, the writing work disappears from the dashboard. Success sends the credit elsewhere. Excess volume sends the cost there too.
Leaders often describe technical writers as the people who produce documentation. That description captures the artifact and misses the job.
The job may include finding the source behind a claim, identifying contradictions, testing a procedure, choosing what not to publish, negotiating with subject-matter experts, designing retrieval paths, recording decisions, managing versions, planning maintenance, and knowing when an answer is too uncertain to release.
Across those roles, people protect the boundary between material that exists and material that deserves institutional trust.
Cheap Generation Can Create an Expensive Organization
A faster drafting tool can make an organization slower.
More drafts create more interruptions for subject-matter experts. Plausible errors take longer to disprove than obvious ones. Weak submissions delay valuable work. Bad documentation becomes support volume. Unowned content becomes maintenance debt.
That is why “human in the loop” is not a cost model.
Which human? How many drafts can they responsibly review? What evidence arrives with the draft? What authority do they have to reject it? Which existing work does the added volume displace?
If the plan is simply for employees to absorb more generated material, the organization does not have an AI workflow.
It has an unfunded labor transfer.
The Repository Put a Price on Attention
arXiv responded by throttling the input.
The universal cap is blunt. It reduces load without distinguishing between a high-volume source of low-value material and a legitimate researcher whose completed work happens to arrive in clusters. Because arXiv also serves as a public record of when research became available, delay can affect visibility, access, and priority.
arXiv describes the cap as a stopgap while it improves moderation tools and procedures.
That tradeoff matters. But so does the boundary the policy establishes: the system does not owe any submitter unlimited moderation labor.
The repository had to impose a visible limit because demand had already exceeded the invisible one—human attention.
Price the Whole Information System
Organizations adopting AI for documentation should stop measuring success at the draft.
The real cost includes:
- Time spent verifying or rejecting generated material
- Interruptions imposed on specialists
- Rework created by missing evidence or context
- Downstream support and correction
- Ownership throughout the document’s useful life
- The high-value work reviewers cannot do while clearing low-value volume
The point is not that every AI-assisted draft is bad, nor that organizations should reject useful automation. The point is that the accounting has to be honest.
If AI genuinely reduces the total work required to produce and maintain trustworthy information, that value should survive measurement across the whole system.
If the apparent savings depend on employees quietly absorbing more review, cleanup, and risk, the tool is not cheap.
The People Were the System
Organizations gravitate toward AI writing because software appears inexpensive beside an employee’s salary.
But the salary was never buying keystrokes alone—it was buying judgment, context, institutional memory, skepticism, coordination, accountability, and the willingness to say that something was not ready.
When leaders do not understand what employees actually do, they calculate automation savings against the wrong job. They replace the visible act of writing, then rediscover the invisible work through delay, errors, support volume, and burnout.
arXiv’s rate limit is a warning from a publishing system that reached that point in public.
The text was cheap.
The system still required people.
AI writing is not cheap. It sends the bill to someone whose work the calculation never counted.