arXiv plans to penalize researchers who submit papers with clear signs of unchecked AI-generated text. The preprint site wants to slow the spread of low-quality, machine-written research.
arXiv will ban authors for one year if their papers show “incontrovertible evidence” that they did not review large language model output. TechCrunch also reported that the policy targets researchers who “let AI do all the work” in scientific papers.
Thomas Dietterich, chair of arXiv’s computer science section, explained the new policy. Dietterich gave examples of clear evidence, such as fake references and leftover LLM comments, like prompts asking users to swap sample data for real experimental numbers.
A one-year ban and stricter future submissions
The penalty is more than just rejecting a paper.
The Verge said that authors who break the rule will be banned from arXiv for a year. After the ban, any new papers must first be accepted by a reputable peer-reviewed venue.
TechCrunch reported the same, adding that future submissions must pass peer review before being posted on arXiv again.
The rule does not ban all use of AI tools. arXiv still allows researchers to use LLMs, but authors must take full responsibility for everything in their papers, no matter how it was created.
Dietterich said that if AI creates inappropriate language, plagiarism, bias, mistakes, wrong references, or misleading claims, the authors are still responsible.
Moderators will need clear evidence
arXiv wants the rule to apply only to clear cases, not to debates about writing style. Dietterich called it a “one-strike” rule. A moderator must document the issue, and a section chair must confirm the evidence before any penalty.
Authors can appeal bans, and the policy only applies when there is clear evidence of unchecked LLM output.
This distinction is important because AI-assisted writing is now common in academia. arXiv’s concern is not about using AI, but about authors submitting work that is so poorly checked that the platform cannot trust the paper.
Dietterich warned that if a paper clearly shows the authors did not check LLM-generated content, “we can’t trust anything in the paper.”
The pressure on preprint platforms is growing
This new penalty builds on earlier actions by arXiv to cut down on low-quality AI-generated papers. arXiv changed its rules so that computer science review articles and position papers can only be posted if they have already been peer reviewed and accepted at a conference or journal.
arXiv explained that large language models made it easy to create review-style content quickly, leading to many submissions that were just annotated bibliographies instead of real research discussions.
arXiv has other rules to prevent low-quality submissions, like making first-time authors get an endorsement from an established researcher. arXiv’s role is sensitive because it publishes research before peer review, but it is now a main way for work to circulate in fields like computer science and mathematics.
A warning to researchers using AI
The policy makes it clear to researchers: using AI to help write or edit a paper is fine, but submitting unchecked machine output is not allowed.
For arXiv, this rule protects trust in a system that relies on authors being responsible. For academics, it is a reminder that while AI can make writing faster, it does not remove accountability.
In preprint research, a fake citation is now more than just a mistake—it could mean losing access to one of science’s key platforms.
