Analysis: scientists who appeared to use LLMs posted 33% more papers on arXiv than those who didn't, as concerns grow over AI slop in scientific publishing
Context & Ripple Effects
Scientific publishing had already drawn a line between AI assistance and authorship: Springer Nature permitted disclosed writing help while declining to credit LLMs as authors a disclosure-based AI-writing policy. Evidence that LLMs had also entered peer review—where a study found AI-written sentences in computer-science reviews LLM-generated language in peer reviews—made research workflows a broader integrity issue, not just an authoring one.
This analysis adds a measurable output gap to that debate. Its importance is less that it proves causation than that apparent LLM use is associated with substantially higher arXiv posting volume, raising the stakes for disclosure and quality controls as preprint production accelerates.
First-order effects
- Scientists who appear to use LLMs have a reported 33% arXiv-posting advantage, while non-users face a potential output disadvantage in a publication channel where volume is visible.
- arXiv readers, moderators, and prospective reviewers must process more material while distinguishing legitimate assisted work from low-quality or improperly attributed AI-generated submissions.
Second-order effects
- Research conferences are likely to tighten or clarify writing and review rules; related coverage says conferences have already moved to restrict LLM use after an influx of generated material conference restrictions on AI-written submissions and reviews.
- The productivity incentive and weaker barriers to drafting can also widen the academic-fraud problem, consistent with later findings that major LLMs can help non-researchers submit fabricated arXiv papers LLM-enabled fabricated-paper submissions.
Third-order effects
- If output gains persist, scholarly publishing may shift from treating AI assistance as an individual disclosure question toward workflow-level provenance, screening, and accountability systems.
- The durable tension is between faster research production and scarce human verification capacity; rules that preserve legitimate assistance without normalizing unverifiable work will shape trust in preprints.
The trend: This is one data point in the shift from generative AI as a writing aid to a productivity layer that forces scientific institutions to redesign quality assurance.