arXiv Bans AI-Generated Slop: What New Moderation Means for Scientific Publishing

## The Ban Hammer Falls

arXiv — the preprint server that hosts most of physics, computer science, and mathematics research — has started banning authors who submit AI-generated content. One moderator described the policy bluntly on social media: "Send the arXiv AI-generated slop, get a yearlong vacation from submissions."

The policy targets papers that are clearly generated by LLMs: formulaic structure, hallucinated references, coherent syntax but nonsensical scientific content. These are not sophisticated frauds — they're volume attacks from people using ChatGPT as a paper mill.

## The Moderation Problem

arXiv receives 20,000+ submissions per month. Human moderators can't read every paper. The current moderation stack:
- **Plagiarism check**: catches copy-paste, doesn't catch AI generation
- **Endorsement system**: requires established authors to endorse newcomers, but this is gamed
- **Automated filters**: new AI-detection tools deployed, accuracy unproven

The fundamental tension: arXiv was designed as a low-barrier preprint server to accelerate science. Requiring heavy moderation contradicts its founding philosophy. But without it, the signal-to-noise ratio degrades to zero.

## What Counts as "AI-Generated Slop"

The policy targets clear-cut cases:
- Papers with no identifiable research contribution
- References to nonexistent papers (LLM hallucinations)
- Generic, repetitive prose that doesn't engage with prior work
- Dozens of near-identical submissions from the same author

Crucially, the policy does NOT target legitimate uses of AI in research: LLM-assisted proofreading, code generation for experiments, or AI methods papers. The distinction: "was AI used as a tool, or as a substitute for thinking?"

## The Broader Implications

arXiv's move is a canary in the coal mine for scientific publishing. If the world's largest preprint server can't filter AI slop, peer-reviewed journals — which already face a reviewer shortage — will drown. The nightmare scenario: researchers spend more time detecting AI fraud than evaluating real science.

This also raises an uncomfortable question about LLM capabilities. If the best LLMs in 2026 are still producing slop that even arXiv's basic filters can detect, what does that say about the gap between AI "writing" and actual intellectual contribution?

For researchers and engineers: the arXiv story is a reminder that tool quality matters. Use AI to assist thinking, not replace it. The detectors are getting better, and the consequences — a yearlong publication ban — are real.

arXiv Bans AI-Generated Slop: What New Moderation Means for Scientific Publishing

// COMMENTS

ON THIS PAGE