Why is AI editing so much heavier in computer science papers, at 17.5% of text versus 6.3% in math?

This explores why computer science leads other fields in how much of its writing shows LLM editing. The corpus doesn't directly answer the 'why', but it documents the gap and gives some clues about it. The core measurement comes from an analysis of nearly a million papers on arXiv, bioRxiv and in Nature journals. It found LLM-modified text rising steadily after ChatGPT's release, reaching 17.5% in computer science and 6.3% in mathematics How much scientific writing has LLMs actually modified?. That study measures how much text was modified. It doesn't test explanations, so anything about causes is inference from nearby findings, not a settled result.

The most suggestive clue is that LLM use in AI-related fields goes beyond papers. Between 6.5% and 16.9% of peer reviews at major AI conferences (ICLR, NeurIPS, EMNLP) show substantial LLM modification, with higher rates among reviewers who were rushed, less confident or less engaged How much peer review text shows signs of LLM modification?. That points to two forces worth considering together. Computer scientists are closest to the tools and most comfortable with them, and CS publishing runs on fast conference deadlines that reward speed. The finding about reviewers suggests time pressure, not just familiarity, predicts who reaches for an LLM.

Incentives point the same way. Across 2.1 million preprints, scientists who adopted LLMs published 23.7–89.3% more papers Does LLM writing assistance change how scientists publish?. In a field where output volume and conference acceptances drive careers, that productivity boost is a strong pull. Within computer science, the LLM share also depends on the kind of paper. Review papers carry a higher share (21.4% vs 14.0%), but ordinary research papers produce about six times as many generated papers by volume Where is most LLM-generated content actually appearing in computer science?. So the field's high average isn't just a niche of survey papers. It runs through everyday research writing.

The less obvious part is why nothing pushes back. Readers with machine-learning expertise can't reliably tell LLM-written abstracts from human ones, and LLM-edited abstracts received the highest clarity ratings Can readers tell LLM abstracts from human ones?. Meanwhile, the link between polished, complex prose and paper quality has reversed among LLM adopters Does LLM writing assistance change how scientists publish?. If polish no longer signals merit and readers can't detect the tool, there's little social cost to using it. That would let adoption spread fastest in the community that already has the tools at hand.

One caution about a popular story. You might expect LLM-assisted reviewers to favour LLM-assisted papers, which would create a feedback loop that speeds up adoption. Across 125,000+ reviews, that apparent favouritism disappears once paper quality is taken into account. LLM papers tend to be weaker, and LLM-assisted reviewers are generally more lenient toward weaker work Do LLM reviewers actually favor LLM-written papers?. So the cycle is real at the level of volume and pressure, but there's no evidence of models rewarding their own writing. For a direct study of why computer science differs from other fields, such as one comparing deadline structures or tool access, the collection doesn't yet have one.