arXiv Imposes Strictest Rules: AI-Generated Papers Banned, All Authors Punished; Terence Tao Endorses
arXiv's computer science chair Thomas Dietterich announced a new policy: papers with clear signs of AI-generated content will lead to a one-year ban for all listed authors, and subsequent submissions must pass peer review. Mathematician Terence Tao supports the move, saying it aligns with the need to prioritize digestion over generation. Princeton University also ended its 133-year honor code due to widespread AI cheating, reinstating proctored exams.
In a decisive move to combat the rising tide of AI-generated academic papers, the arXiv preprint repository has announced its strictest policy to date: authors found submitting content with clear, unchecked AI contributions face a one-year ban from the platform, with all co-authors held collectively responsible. The new rules, unveiled by Thomas Dietterich, chair of arXiv’s computer science section, signal a firm stance against the misuse of generative AI in scientific publishing.
The policy targets specific red flags that indicate authors did not carefully review AI-generated material. These include hallucinated references to nonexistent papers, residual LLM commentary such as “This is a 200-word summary, do you need modification?” left in the manuscript, and unfilled placeholders like “Please insert experimental data here.” Dietterich emphasized that such errors render the entire paper untrustworthy. Once confirmed, the penalty is a one-year suspension, after which any new submission must first pass peer review at a legitimate journal before reappearing on arXiv. Appeals are allowed, but the process is designed to be “one-strike.”
Prominent mathematician Terence Tao, an outspoken advocate for integrating AI in research, has endorsed the policy. In a series of posts on Mathstodon, Tao compared the new rules to a framework he had outlined earlier, which includes clearly defining acceptable AI assistance, reducing emphasis on rapid publication, designing new challenges for heavy AI users, and transparently communicating project goals. He noted that arXiv’s policy aligns with the first two points, especially the need to rebalance the academic ecosystem toward “digesting” research rather than “producing” it. “In an era where generating a paper is far easier than digesting one, any effort that tilts the traditional balance back toward digestion is welcome,” Tao wrote. He also acknowledged that the policy cannot address the systemic “publish or perish” culture, but called it a necessary step.
The severity of AI-generated paper flooding is underscored by a 2025 study in Nature Human Behaviour, which found that about one-fifth of computer science papers showed clear signs of LLM modification, with a sharp increase after ChatGPT’s release. arXiv now receives hundreds of AI-generated survey papers each month—content historically written by senior experts. Academic conferences are also affected: a GPTZero analysis of NeurIPS 2025 papers identified hundreds of hallucinated citations in at least 53 accepted papers. The problem extends beyond publishing; Princeton University recently abolished its 133-year-old honor system after a survey revealed over 27% of seniors admitted to using AI to cheat on exams, prompting a return to proctored testing.
Tao further suggested that alternative platforms like viXra could serve as archives for AI-heavy content, but cautioned that such repositories should remain outside the formal citation chain. The overall goal, he argued, is to prevent a feedback loop where low-quality AI-generated papers contaminate training data, leading to even worse papers—a self-imposed “lock” on scientific progress reminiscent of the alien “sophon” in Liu Cixin’s Three-Body Problem. As Dietterich put it, “Signing means taking responsibility, no matter how the content is generated.”