ArXiv implements one-year author ban for unchecked AI-generated content

ArXiv, the widely used preprint repository that has anchored scientific communication in computer science and mathematics for over two decades, is escalating enforcement against inadequately vetted AI-generated submissions. Thomas Dietterich, chair of the computer science section, disclosed Thursday that papers containing direct evidence of unverified LLM generation will trigger author sanctions.

The policy targets concrete red flags: fabricated references, visible conversational artifacts with language models, or other unmistakable signs that authors failed to validate machine-generated content. This represents a precision enforcement approach—the prohibition is not on LLM use per se, but on the abdication of authorial responsibility for accuracy.

Penalties for confirmed violations are structured in two tiers. First-time offenders face a one-year ban from posting to ArXiv. Any subsequent submissions must first achieve acceptance at a peer-reviewed external venue, establishing a higher external validation burden before ArXiv reentry. Before imposing penalties, moderators must flag concerns and section chairs must independently verify the evidence, with appeal mechanisms available to accused authors.

The enforcement framework reflects ArXiv's broader effort to manage research quality degradation. The repository has already implemented endorsement requirements for first-time posters and is transitioning from Cornell's institutional hosting to independent nonprofit status—a structural change intended to enable sustained funding for content moderation and abuse prevention as AI-generated submissions proliferate.

Separately, peer-reviewed literature has documented rising rates of fabricated citations in biomedical research, a trend researchers have linked to LLM hallucinations generating seemingly plausible but entirely fictional references.