arXiv is drawing a harder line around AI-assisted papers: use the tools if you want, but unchecked hallucinations can now cost authors access to the platform.
The warning is simple, and it lands hard because arXiv sits at the center of modern scientific publishing. A submission that carries obvious evidence of unreviewed LLM output, including fake citations, stray model instructions or invented results, can now trigger a one-year ban from arXiv and a tougher path back for future papers.
That does not mean arXiv is banning AI-assisted writing. It means the platform is treating unchecked AI output as an authorship failure, not a technology problem. If a researcher uses a model to draft, translate, summarize or format a manuscript, the responsibility still sits with the human names on the paper. That has always been the practical rule. Now the penalty is becoming much harder to ignore.
According to a May 15 Reddit post reproducing a public thread from Thomas G. Dietterich, an arXiv moderator for cs.LG, arXiv has clarified the consequences for submissions that contain incontrovertible evidence authors did not check LLM-generated material. Dietterich said the penalty is a one-year ban, followed by a requirement that later arXiv submissions first be accepted by a reputable peer-reviewed venue.
The examples matter because they are not subtle. Hallucinated references are not a disagreement over interpretation. They are papers that do not exist, titles that sound plausible, journals that were never involved and bibliographies that only look scholarly at a glance. LLM meta-comments are even more direct, the kind of leftover instruction text that tells a reader the manuscript passed through automation without basic human review.
Illustrative fake results sit in the same category. A table that says the data should be replaced with real experimental numbers is not a formatting issue. It tells moderators that the authors may not have verified the evidence supporting the paper. Once that happens, the problem is no longer one paragraph. It is the credibility of the entire submission.
This is why the arXiv move matters beyond one platform. Preprint servers were built for speed. Researchers use them to share work before conferences, journals and peer reviewers have finished their slower processes. That speed is valuable, especially in computer science, machine learning, physics and mathematics, where waiting months can mean falling behind the conversation. But speed only works when authors bring discipline with them.
LLMs change that balance. They make it easier to produce polished academic language, generate summaries, expand related work sections and format references. They also make it easier to create a manuscript that looks complete before it has been checked. The danger is not that researchers use AI. The danger is that the fluency of the output hides the absence of verification.
The citation numbers are no longer theoretical
The timing is not accidental. A May 8 arXiv preprint auditing 111 million references across 2.5 million papers and preprints in arXiv, bioRxiv, SSRN and PubMed Central estimated 146,932 hallucinated citations in 2025 alone. Nature reported on May 14 that SSRN had the highest rate among the major repositories studied, at nearly 2 percent, almost five times higher than any other major repository in the audit.
Those figures do not prove every bad citation came from an LLM. Citation errors existed long before ChatGPT. Authors mistype names, copy broken bibliographies and cite papers they have not read closely enough. But the sharp rise after widespread LLM adoption gives moderators a practical problem: the old safeguards were not designed for automated systems that can produce convincing fake scholarly scaffolding at scale.
For researchers, the compliance burden is now part of the workflow. It is no longer enough to say a model was used only for language assistance. Every reference needs to resolve. Every result needs to trace back to real data. Every sentence that sounds like a claim needs support. That is basic scholarship, but AI tools make skipping those steps easier and more tempting.
The burden will fall hardest on fast-moving AI labs, early-career researchers and independent authors who rely on arXiv visibility before journal acceptance. A one-year ban is not just an inconvenience. For some authors, it can interrupt conference timing, hiring signals, collaboration cycles and the public record of their work. That is exactly why the penalty has teeth.
There is also a market implication for the AI tooling industry. Reference managers, writing assistants and research copilots will need stronger verification features, not just cleaner prose. A product that helps draft a related work section but cannot prove the citations exist is now creating risk for its users. The next useful layer in academic AI may be less about generation and more about audit trails.
arXiv is not trying to solve hallucination across science by itself. It is setting a boundary at the point where scientific infrastructure meets automated writing. Authors can still use powerful tools, but the platform is making one thing plain: if the machine invents, and the human does not check, the human owns the failure.
What to watch next is whether conferences, journals and funders follow the same approach. Soft disclosure rules are useful, but they do not change behavior on their own. Clear penalties might. The researchers who adapt fastest will not be the ones who stop using AI. They will be the ones who build verification into every step before the paper ever reaches submit.
Also read: Tech layoffs are changing the startup hiring market in 2026 • LTX Director turns AI video into an editable indie workflow. • TurboQuant gives AI startups a useful reminder about inference costs
