Large language models are being utilized carelessly in scientific articles, and ArXiv, a popular open repository for preprint research, is taking more action.
ArXiv (pronounced “archive”) has emerged as one of the primary channels for the dissemination of research in disciplines like computer science and mathematics, despite the fact that articles are uploaded to the site prior to peer review. Additionally, the site has developed into a source of information on scientific research trends.
ArXiv has already taken action to counteract an increasing number of subpar AI-generated publications, such as mandating that new posters obtain an endorsement from a reputable author. Additionally, the organization is becoming an independent nonprofit after being housed by Cornell for over 20 years, which should enable it to raise more funds to address problems like AI slop.
The chair of arXiv’s computer science section, Thomas Dietterich, recently tweeted on Thursday that “if a submission provides clear evidence that the authors did not examine the outcomes of LLM creation, this indicates we can not trust anything in the work.”
“Hallucinated references” and remarks to or from the LLM are examples of such unquestionable evidence, according to Dietterich. The authors of a publication will be subject to “a 1-year ban from arXiv followed by the condition that subsequent arXiv submissions must first be accepted by a respectable peer-reviewed venue” if such proof is discovered.
It should be noted that this is an urge that authors assume “full responsibility” for the material, “irrespective of how the contents are generated,” rather than a complete ban on the use of LLMs. Therefore, researchers are still accountable if they copy and paste “inappropriate language, plagiarized content, prejudiced content, errors, blunders, wrong references, or deceptive content” straight from an LLM.
This will be a “one-strike” rule, Dietterich told 404 Media, but before the penalty is applied, moderators must report the problem and section chairs must verify the evidence. The verdict will also be appealable by the authors.
