arXiv might not survive AI slop onslaught, warn mathematicians

The preprint database arXiv has been forced to limit paper submissions to stem a vast and quickening flood of AI-generated papers. Academics tell New Scientist that the massive influx of AI-generated papers is threatening to overwhelm the database and render it useless. ArXiv offers a place for academics to host papers that haven’t yet been peer-reviewed or published in reputable journals. This gives colleagues a chance to keep up to date with what’s ahead in fields as diverse as physics, astronomy and economics. But AI’s dramatic growth in mathematical prowess and its ability to automatically generate papers at speed with minimal human effort mean that the arXiv website is being flooded with an increasing number of papers. Advertisement In September 2024, 20,569 papers were submitted to arXiv. The following September, that number had grown to 26,646. But in a dramatic surge, 40,363 papers were submitted in September 2026. Aleksandr Logunov at the Massachusetts Institute of Technology warns that the problem is serious. “Right now, journals are flooded with submissions of AI articles,” he says. Kevin Buzzard at Imperial College London says it has also become harder to submit genuine papers because arXiv moderators are overwhelmed, and papers can now disappear into the review queue for weeks at a time. The number of published papers is so large that Buzzard can’t manually read them every day. Ironically, he now has to use AI to tell him which papers to focus on. “ArXiv is a serious problem, and it might not survive the onslaught,” says Buzzard. “Journals are also under more and more pressure. The issue is that slop papers do not explain stuff well. It’s not just about correctness; it’s about human understanding, and AI is sometimes very poor at explaining.” With access to powerful AI tools, non-professional mathematicians are now able to solve reasonably difficult problems and automatically generate papers outlining the results that can then be posted to arXiv. But such papers are often criticised by mathematicians. For instance, Chaim Goodman-Strauss at the University of Arkansas described one paper this week – which included a valid solution to a long-standing problem – as “basically a piece of slop”. Terence Tao at the University of California, Los Angeles, recently complained to New Scientist that the way in which discoveries are being made and “dumped” on the mathematical community may actually harm the field. In response to the problem, arXiv spokesperson Kat Boboris wrote in a blog post that there are now limits on submitters. They are permitted to upload just two papers each calendar month, and can have three total active submissions at any given time. ArXiv policy has long permitted authors to use AI, as long as it is disclosed. But Boboris wrote that arXiv has witnessed “an increase in thin papers of narrow scope” and “low value” research that has forced it to find new measures to uphold quality. ArXiv didn’t respond to a request for an interview, but the blog post states that, prior to the appearance of AI models that can generate papers, there was a practical limit to submissions, and that moderators would limit submitters who seemed to go beyond that. Other projects are under way, aiming to adapt to the new reality of AI-enabled research. Among them is new site called Hexagon. While arXiv offers a place to host “preprints” that aren’t yet peer-reviewed and published, Hexagon is a place for “pre-preprints” says Buzzard – somewhere to host AI-generated papers that mathematicians may not have fully understood yet. [Image text:] ar Giv ar Xiv