ArXiv, a leading platform for preprint academic research, is implementing new measures to decrease the influx of papers containing AI errors. Papers that display “incontrovertible evidence that the authors did not verify the results of LLM generation,” including fabricated references or “meta-comments” from an LLM, will result in a one-year ban from ArXiv, as stated by Thomas Dietterich, the section chair for computer science at ArXiv. Additionally, future submissions to ArXiv must be accepted by a “reputable peer-reviewed venue.”
Dietterich shared a message on X, emphasizing that according to the Code of Conduct, each author is responsible for the content of their paper, regardless of how it was generated. If inappropriate, plagiarized, biased, or erroneous content produced by generative AI tools is included in scientific works, it is the authors’ responsibility. Papers containing such AI-generated mistakes will be subject to new penalties, including a one-year ArXiv ban and the requirement that future submissions must first be peer-reviewed and accepted elsewhere. Examples of “incontrovertible evidence” mentioned include fabricated references and meta-comments like, “here is a 200 word summary; would you like me to make any changes?”
Dietterich told 404Media that ban decisions can be appealed, and clarified the policy applies only when incontrovertible evidence is present. The internal process requires a moderator to first identify the issue, followed by confirmation from the Section Chair before penalties are enforced.
Last year, ArXiv revised its policies to limit AI errors by only allowing computer science review articles and position papers if they had undergone peer review and were accepted by a conference or journal. ArXiv noted the rise of large language models has made the creation of this content easy, resulting in many submissions being annotated bibliographies without substantial discussion on research issues.