Fighting AI slop in science papers (www.unite.ai)

In a significant move to combat the deluge of low-quality AI-generated scientific papers, preprint server arXiv announced a new policy limiting authors to two submissions per month. This decision follows a striking increase in submissions, from 9,869 in September 2016 to 40,363 in September 2026, raising concerns over the rising volume of thin and poorly composed papers, often facilitated by AI tools. ArXiv's moderators report an alarming uptick in "scientific slop," characterized by shallow arguments and disconnected evidence within submissions. To address these issues, arXiv has previously banned unchecked AI content and required additional peer support for survey and position papers. Amid these challenges, researchers from Seoul National University and the University of Minnesota have developed a framework called SciSlopHarness to measure and repair "scientific slop" in AI-generated papers. Their approach focuses on the connections among a paper's claims, arguments, and evidence, rather than just the text itself. The new benchmark, SciSlopBench, demonstrated an 85.9% accuracy in distinguishing AI-generated content from human-written papers. This innovative method not only aims to enhance the quality of scientific submissions but also sets a precedent for improving the evaluation of academic work in an AI-driven landscape.