arXiv Imposes Year-Long Ban for Researchers Who Let AI ‘Hallucinate’ Their Papers
Table of Contents
The High Cost of Copy-Paste
For decades, arXiv has served as the digital plumbing of the scientific world. By allowing researchers in physics, mathematics, and computer science to share preprints before they hit the grueling gauntlet of formal peer review, the platform has accelerated the pace of discovery. But the rise of large language models (LLMs) has introduced a new, volatile variable into the mix: “AI slop.”
The repository is now drawing a hard line. Thomas Dietterich, chair of arXiv’s computer science section, announced Thursday that authors who submit papers containing incontrovertible evidence of unvetted AI generation will face a one-year ban from the platform. Once that ban expires, those authors will be required to prove their credibility by getting subsequent submissions accepted by a reputable peer-reviewed venue before they can post to arXiv again.
This isn’t a ban on AI tools themselves—which are increasingly integrated into data analysis and drafting—but rather a crackdown on negligence. The core of the issue is accountability. According to Dietterich, if a paper contains evidence that the authors didn’t even bother to check the results of an LLM’s generation, the integrity of the entire document is compromised.
Spotting the ‘Hallucinations’
Identifying AI-generated fraud isn’t always a matter of sophisticated detection software; sometimes, it’s as simple as reading the text. Dietterich noted that “hallucinated references”—citations of papers that simply do not exist—are a primary red flag. Even more damning are the “smoking guns”: instances where researchers accidentally leave in prompts or responses from the AI, such as “As an AI language model, I cannot…” or formatting markers unique to ChatGPT and Claude.
When these markers appear, it suggests a level of carelessness that transcends a mere mistake. It signals that the author essentially outsourced the intellectual labor of the research to a machine and bypassed the critical step of human verification. Under the new “one-strike” policy, such evidence leads directly to the suspension of posting privileges.
To prevent arbitrary censorship, the process involves a system of checks and balances. Moderators must first flag the suspected content, and section chairs must confirm the evidence before the penalty is applied. Authors will also have a window to appeal the decision, acknowledging that the line between “AI-assisted” and “AI-generated” can sometimes be blurry.
A Growing Crisis of Trust
The move comes as the broader scientific community grapples with a surge in fabricated data. Recent studies in biomedical research have shown a disturbing uptick in fabricated citations, a trend largely attributed to the ease with which LLMs can generate plausible-sounding but entirely fake academic references. This phenomenon isn’t limited to the fringes of academia; it has cropped up across various disciplines, threatening the fundamental trust required for scientific progress.
arXiv is also undergoing a structural shift to handle these challenges. After more than 20 years of being hosted by Cornell University, the organization is transitioning into an independent nonprofit. This move is intended to diversify its funding and provide the resources necessary to build more robust moderation tools to combat the flood of low-quality submissions.
The repository has already implemented some barriers to entry, such as requiring first-time posters to be endorsed by an established author. However, the new ban suggests that endorsement is no longer enough. As the barrier to producing academic-looking text drops to zero, the value of the human editor—and the risk of the lazy researcher—has never been higher.
