ArXiv's one-year ban on AI slop submitters is the right policy, but it is also an admission of failure. The preprint server's decision to penalize researchers who submit hallucinated citations and unedited LLM meta-comments reveals how thoroughly AI-generated content has already contaminated the scientific record—and how few tools exist to catch it before publication.
Thomas Dietterich, chair of ArXiv's computer science section, announced the enforcement on social media Thursday, specifying that submissions containing "incontrovertible evidence" of unchecked AI generation would trigger the ban. The examples he cited are damning: fabricated references pointing to papers that do not exist, direct messages from language models left in the text ("here is a 200 word summary; would you like me to make any changes?"), and prompts asking authors to fill in real experimental data after the AI had supplied illustrative figures. Dietterich's conclusion was stark: "If a submission contains incontrovertible evidence that the authors did not check the results of LLM generation, this means we can't trust anything in the paper."
The penalty—a year-long submission embargo followed by a permanent requirement to publish in peer-reviewed venues first—sends a clear message. But it also underscores the scale of the problem. ArXiv is not catching AI slop; it is reacting to submissions so careless that obvious artifacts went undetected. The contamination runs deeper than a few bad actors. Researchers are submitting AI outputs without basic review, and the evidence is sometimes literally embedded in the document as LLM instructions to the author.
Critics will argue that the ban is too blunt an instrument. Legitimate researchers use AI to improve writing, check grammar, or even generate initial drafts—activities that are not inherently problematic. A blanket prohibition on AI-assisted work misidentifies the culprit. The policy itself acknowledges this distinction: the target is not AI use but the absence of human oversight. Yet enforcement remains murky. "Incontrovertible evidence" requires obvious artifacts that a diligent reviewer would catch. Subtle cases—plausible-sounding but fabricated studies, subtly biased literature reviews—will still pass through.
The deeper concern is what this ban reveals about peer review itself. ArXiv's permanent peer-review requirement for banned researchers implies that traditional journal gatekeeping can catch what preprint servers cannot. But fake citations and LLM artifacts have appeared in peer-reviewed literature as well. The contamination is not a preprint problem; it is a systemic one. The volume of submissions has overwhelmed human reviewers at journals and conferences alike, creating conditions where AI slop can hide in plain sight.
What ArXiv has done is necessary. It has drawn a line and established consequences. But the real story is the underlying crisis: researchers are flooding academic infrastructure with AI-generated work faster than institutions can adapt their defenses. The ban buys time. It does not solve the problem.