arXiv、AI生成のずさんな投稿者を1年間投稿禁止に Send the arXiv AI-generated slop, get a yearlong vacation from submissions

arXiv, the preprint server that has become indispensable infrastructure for physics, computer science, and mathematics research, is sharpening its response to a rising tide of AI-generated submissions. The site announced it will ban submitters who upload papers riddled with AI hallucinations or otherwise obvious large language model slop for up to a year.

Unlike peer-reviewed journals, arXiv serves as an open repository where researchers share findings quickly, often before formal review. That openness has made it a vital channel for scientific communication, but it has also left the platform vulnerable to the flood of plausible-sounding yet hollow manuscripts produced by tools like ChatGPT. Moderators have reported a sharp uptick in submissions featuring fabricated citations, nonexistent equations, and factually incorrect claims dressed up in confident academic prose.

The problem has been especially acute in review-style papers, where LLMs can generate superficially coherent surveys with little underlying expertise. Earlier this year, arXiv tightened rules around review articles, requiring authors to first secure endorsements from recognized moderators. The new yearlong ban escalates enforcement further, signaling that repeat or egregious offenders will face meaningful consequences rather than mere rejection of a single paper.

Notably, arXiv is not banning AI-assisted writing outright. The policy appears aimed at submitters who outsource the substance of their work to a model without verifying its output, rather than those who use LLMs for legitimate editing or drafting support. That distinction mirrors the stance taken by major journals: Nature, Science, and publishers like Elsevier and Springer Nature have all stated that AI tools cannot be listed as authors and that any use must be disclosed transparently.

The broader ecosystem response reflects growing anxiety about the integrity of scientific literature in the LLM era. Tools designed to detect AI-generated text have proven unreliable, with false positives that can unfairly tarnish legitimate researchers, particularly non-native English speakers whose writing styles can trigger detectors. How arXiv will operationalize its judgments about what constitutes AI slop, and whether it will publish clear criteria, remains an open question. Transparency around enforcement will likely determine whether the policy succeeds in deterring bad actors without chilling legitimate use.

The stakes extend beyond arXiv itself. Preprint servers have become a core part of how modern science is communicated, with papers frequently cited, discussed, and built upon long before any journal version appears. If the signal-to-noise ratio degrades too far, the value of the entire preprint ecosystem could erode. Other repositories, including bioRxiv and SSRN, may be watching arXiv's experiment closely as they grapple with similar pressures. For now, arXiv's move suggests that even the most permissive corners of scientific publishing are drawing harder lines against unchecked generative AI output.