arXiv AI-Generated Slop: Submit Trash, Get a 12-Month Vacation from Research

arXiv just dropped the hammer on arXiv ai-generated slop ban, and the penalty is a mandatory 12-month vacation from the platform. Starting this week, the world’s most popular preprint repository is using updated detection algorithms to flag papers that are essentially raw outputs from models like GPT-4o or Claude 3.5 Sonnet. It is a move to save the site from being buried under thousands of low-effort, hallucinated research papers that offer zero value. If you try to cheat the system with unedited AI text, you are out for a year.

The New 365-Day Suspension Policy

arXiv updated its submission guidelines on May 12, 2026, introducing a strict ‘one-strike’ rule for what they call ‘automated low-utility content.’ If their automated screening system, now powered by a custom version of GPT-4o-level detection, flags your paper as 90% or more AI-generated without significant human revision, your account is toast for 365 days. I think this is long overdue. The repository saw a 22% spike in submissions over the last 18 months, much of it being ‘slop’—papers with fake citations and circular logic that waste the time of every serious researcher. This isn’t just a slap on the wrist; it’s a career-stalling move for academics who rely on arXiv for visibility. If you are a PhD student and you get banned, good luck explaining that 12-month gap in your publication record to your advisor.

Why a year-long ban?

The 12-month duration is intentional. It covers a full academic cycle. arXiv administrators stated that the manual review process for appeals costs the platform roughly $150 per paper in labor. By enforcing a long ban, they discourage ‘carpet bombing’ the site with multiple AI drafts to see what sticks. If you lose access, you lose the ability to claim priority on new discoveries for an entire year.

How arXiv-Guard Detects the Slop

The tech behind the ban is a proprietary tool colloquially called ‘arXiv-Guard.’ It doesn’t just look for ‘AI-style’ writing; it cross-references every single citation against real-world databases like CrossRef and PubMed. Last month, analysts found that nearly 15% of rejected papers contained at least one completely fabricated citation—a classic hallmark of Gemini 2.0 or GPT-4 hallucinations. The system also analyzes the ‘perplexity’ of the text. If your paper reads with the uniform flatness of a standard LLM output, it triggers a manual review. I’ve seen some of these ‘slop’ papers, and they are embarrassing. They use phrases like ‘it is important to consider’ every three sentences. arXiv isn’t banning AI as a tool, but they are banning people who use it to bypass the actual work of thinking.

The Hallucination Filter

The most effective part of the new system is the citation checker. arXiv-Guard flags any paper where more than 5% of the bibliography cannot be verified. Since LLMs love to invent papers that sound plausible but don’t exist, this is a death sentence for lazy submitters. It’s a binary check: either the DOI exists, or it doesn’t.

The Real Cost of Academic Noise

Running arXiv isn’t free. While it costs users $0 to submit, the Cornell-hosted service operates on a budget of roughly $3.5 million per year, funded by member institutions and donors. When thousands of AI-generated papers flood the system, it drives up storage costs and, more importantly, breaks the ‘moderation’ layer. Every paper on arXiv is supposed to be skimmed by a human moderator before it goes live. In 2025, moderators were overwhelmed, leading to a 4-day delay in postings. This new ban is designed to clear the pipes. I support this because the signal-to-noise ratio was becoming unbearable. We were reaching a point where searching for ‘Large Language Models’ on arXiv returned 500 papers a day, half of which were absolute garbage written by the models themselves.

Moderator Burnout

Volunteer moderators are quitting because they are tired of reading AI-generated gibberish. By implementing the 12-month ban, arXiv is protecting its human capital. They are prioritizing the 180,000 legitimate researchers over the few thousand scammers trying to pad their resumes with AI-generated volume.

What This Means for Your Research Workflow

Don’t panic if you use Claude 3.5 to help clean up your grammar or summarize your own findings. arXiv isn’t going on a witch hunt for people using AI as a sophisticated spell-checker. They are looking for ‘slop’—content where the AI did the heavy lifting of the research and the writing. If you wrote the code, performed the experiments, and typed the majority of the analysis, you are safe. However, I recommend keeping your drafts and ‘git’ history. If you do get flagged by a false positive, you’ll need that evidence to appeal. The ban is a nuclear option, and while arXiv claims a 99.8% accuracy rate for their detection, that 0.2% error rate could ruin a researcher’s year. Be smart: use AI to polish, not to create.

Keep Your Version History

Always use version control like GitHub or Overleaf’s history feature. If arXiv-Guard flags your paper, showing a clear progression of human writing over several weeks is the only way to get that 12-month ban overturned. Without a paper trail, you are at the mercy of the algorithm.

⭐ Pro Tips

• Run your final draft through a citation checker like Zotero to ensure every DOI is valid before submitting to arXiv.
• Avoid using standard LLM filler phrases like ‘In summary,’ or ‘This underscores the importance’ which trigger detection flags.
• Keep a local copy of your raw data and experiment logs in case you need to prove the research actually happened.

Frequently Asked Questions

Final Thoughts

arXiv’s 12-month ban on AI slop is a brutal but necessary evolution. The platform was built on trust and human curation, two things that ‘slop’ fundamentally destroys. If you’re a serious researcher, this change actually helps you by clearing the noise from your field. If you’re looking for a shortcut to a longer publication list, find a different hobby. My advice? Write your own papers, check your own citations, and treat AI as a tool, not a ghostwriter.