Where is most LLM-generated content actually appearing in computer science?

The authors test the rationale for arXiv's October 31, 2025 policy, which bars unpublished review, survey and position papers from its CS servers. arXiv cited rising LLM-generated content in reviews but gave no data. Here, review papers carry a higher share of LLM-generated content: 21.4% against 14.0% for non-review papers, as adjusted cohort-level estimates over three years. Counted by paper, the volume sits mostly outside reviews. For 2025 the authors estimate 4,783 LLM-generated review papers against 26,801 non-review papers, almost six times as many, out of about 12K review papers and 141K non-review papers. Using the commercial detector Pangram, they put 2025 shares at 43.3% for reviews and 23.3% for non-review papers.

The estimates are population-level by design; the authors do not claim to label individual papers. A classifier first separates review from regular papers, scoring 92.0% F1 on 200 hand-annotated CS papers, half selected for review keywords. Two detectors then estimate the LLM-generated fraction of each group: the Alpha estimator (Liang et al., 2024a), which relies on adjective occurrence and needed a type-specific Rogan-Gladen correction for different false-positive rates, and Pangram, which the authors report produced no false positives. The main pipeline uses titles and abstracts, the same data arXiv moderators saw. The authors add that "review" is not a neutral category, since the annotation follows a "typical" CS view rather than each subfield's norms.

The temporal result sharpens its contrast with How fast did LLM writing adoption actually spread?. That note reports a surge followed by a plateau across four public-facing domains; this excerpt describes "an accelerating trend since 2023" across arXiv's CS, Physics, Mathematics and Statistics. The corpora and estimators differ, so neither settles the other, but they disagree on whether adoption has leveled off. On policy, Does banning LLM use in peer review change review outcomes? measures what a restriction does to review outcomes, whereas this paper measures where generated text sits and finds reviews a minority of it. On detection, Why did AI article share stop growing after 2025? raises the open question of whether a detector-based share reflects use or what the detector can see; this paper addresses that with a correction rather than resolving it.

The excerpt does not establish that generated papers are weaker; it measures how much generated text appears, not its quality. It contains no moderation data, so the claim that the ban "makes little sense" for saving moderator time is an inference from paper counts. The counts also inherit detector uncertainty: the Alpha estimate may shift with the generating model, the estimation data or human post-editing, and the classifier's borderline cases, such as tutorials and hybrid papers, stay ambiguous. The supported implication is narrower than a verdict on the policy: a review-only restriction would miss most of the generated volume, at roughly the magnitude these estimates give. Whether it should be replaced depends on moderation evidence this excerpt does not include.