When Writing Gets Cheap, Reading Becomes the Job
AI made producing work cheap, but reading it still costs human attention. If your output doubles and your reviewers' trust drops, you only moved work onto someone else's desk.
This week, arXiv told every researcher in the world: two papers a month. That's it.
When one of the most open places in science starts rationing submissions, something has changed. And I think it has already reached your pull request queue too.
What happened
On October 1, arXiv, the free preprint server where most AI and physics research shows up first, updated its rate limit policy. Each author can now submit at most two papers per calendar month, with no more than three active submissions at any time.
The numbers behind it are wild. arXiv received 40,363 submissions in September 2026, almost double the 20,569 from September 2024. The cs.AI category alone grew more than six times in two years. That one month also produced almost 9,000 support tickets for staff and volunteer moderators.
arXiv was clear about the cause. AI tools made writing papers much easier, and they saw more "thin papers of narrow scope" and "salami" papers, where one piece of work gets sliced into several smaller ones. In their words, "a relatively small proportion of authors are submitting a large number of low-quality papers and consuming a disproportionate fraction of the moderators' time."
Not everyone agrees with the fix. Scott Kominers from Harvard Business School argued arXiv should economise on moderation, not on scholarship. Others liked it. Mathematician Alex Kontorovich put it simply: "I'm not going to read your 42 papers this month. Pick two, your very best two."
My observation
This is not a science story. It's a software story.
I see the same shift in my own work with AI coding tools. Writing code got cheap. Reading it did not. A pull request that used to take a day to write can now take an hour, but a careful review still takes a human the same focused time. Sometimes more, because the author understands the code less.
So the bottleneck moved. It's no longer the keyboard. It's the attention of the people around you.
When producing becomes cheap, attention becomes the expensive part. Respect it, or someone will start rationing it.
arXiv learned this at scale. Their moderators are volunteers with limited hours, and a few people producing a lot were eating most of that time. Replace "moderators" with "senior engineers" and "papers" with "pull requests", and you have a very familiar picture.
What to expect next
I think more open systems will add limits. Conference reviews, open source maintainers, hiring pipelines, internal review queues. Anything that relies on a human reading what others submit is under the same pressure.
And inside companies, I expect the value of good judgment to go up. Not the person who ships the most. The person whose work is worth reading.
The learning
Volume was never the goal. It was a proxy, and AI broke the proxy. If your output doubles but your reviewers' trust drops, you didn't get more productive. You moved your work onto someone else's desk.
What can we do next
- Curate before you submit. Before opening a pull request, ask: would I want to review this? Cut what isn't needed.
- Don't salami your work. Small PRs are good. Ten PRs with no story connecting them are not. Group changes by intent and explain the intent.
- Own what the AI wrote. If you can't explain a line in review, you are not ready to ask for one.
- Managers, measure differently. Count fewer things like PRs or tickets closed. Look at rework, review time and how often a change needed a second round.
- Protect reviewer time on purpose. Agree as a team how many open PRs one person can have at once. arXiv chose three. Your number can be different, but have one.
Writing is now the easy part. Making your work worth someone's time is the job.
Sources: arXiv blog: Updated Rate Limit Policy, Times Higher Education: arXiv's new preprint submission cap divides opinion
