Did AI use in science level off after the ChatGPT surge, or is some of that plateau just a limit of measurement?
Did LLM use in scientific writing plateau after the initial surge?
This explores whether LLM-assisted writing in scientific papers kept climbing after ChatGPT's launch or levelled off, and how far we can trust the measurements that answer that question.
This explores whether LLM use in scientific writing kept rising after the post-ChatGPT surge or flattened out, and whether the numbers can tell the difference. The short answer from the corpus: the surge is well documented, the evidence on a plateau is mixed, and the plateau that does show up may partly be a limit of the measuring tool.
The surge itself is clear. A study of nearly a million papers from arXiv, bioRxiv and Nature journals found estimated LLM-modified content rising sharply after ChatGPT's release. It reached about 17.5% in computer science and 6.3% in mathematics, so adoption varies a lot by field How much scientific writing has LLMs actually modified?. That study reports steady growth, not flattening. The clearest plateau comes from outside science. Across consumer complaints, corporate press releases, job postings and UN documents, LLM-assisted text jumped after ChatGPT and then settled at roughly 10–24% by 2024, with smaller and younger organizations adopting fastest How fast did LLM writing adoption actually spread?. It's tempting to assume science followed the same curve, but these sources don't show that directly.
The bigger twist is that a plateau may not be a real plateau. These estimates come from statistical fingerprints of model-written text, and a flat line in 2024 fits two very different stories. Either people stopped adopting, or newer models write text the detector can no longer pick out Is the 2024 LLM writing plateau real saturation or measurement artifact?. The framework has no accuracy figures for newer model output, so it can't separate the two. Human readers can't settle it either. People with ML expertise couldn't reliably tell LLM abstracts from human ones, and they preferred LLM-edited abstracts for clarity Can readers tell LLM abstracts from human ones?. As the text gets better, it gets harder to count.
The headline percentage may also matter less than what the adoption is doing to science. Across 2.1 million preprints, scientists who adopted LLMs published 23.7–89.3% more papers. Meanwhile, complex, polished prose stopped being a sign of a stronger paper Does LLM writing assistance change how scientists publish?. Peer review shows a related pattern. LLM-assisted papers cluster among weaker submissions, and an apparent bias of LLM-assisted reviewers toward LLM papers disappears once paper quality is taken into account Do LLM reviewers actually favor LLM-written papers?. So whether or not usage has levelled off, an old shortcut is gone: fluent writing no longer tells you much about the science underneath. That makes the plateau question less important than the question of how reviewers and readers now judge quality.
Sources 6 notes
Analysis of 950,965 papers from arXiv, bioRxiv, and Nature journals found steady growth in estimated LLM modification, peaking at 17.5% in computer science and 6.3% in mathematics, with sharp acceleration after ChatGPT's release.
Across consumer complaints, press releases, job postings, and UN documents, LLM-assisted text rose sharply in the months after ChatGPT's launch and stabilized at roughly 10–24% by 2024. Smaller and younger organizations adopted faster than larger ones.
A plateau in LLM-assisted writing estimates could reflect either genuine adoption saturation or improved model subtlety that evades detection. The framework lacks model-capability data and accuracy figures on newer text to separate these readings.
Readers with ML expertise struggle to identify LLM-generated content reliably, tending to assume human involvement across all abstract types. However, LLM-edited abstracts received highest clarity ratings and were preferred 55% of the time when authorship was disclosed.
Across 2.1M preprints, scientists using LLMs showed 23.7–89.3% higher publication rates. Simultaneously, the correlation between complex prose and paper quality reversed, suggesting polish no longer reliably indicates scientific merit.
Show all 6 sources
Across 125,000+ reviews, the apparent favoritism of LLM-assisted reviewers toward LLM papers disappears once paper quality is held constant. LLM papers cluster among weaker submissions, creating a spurious interaction driven by LLM reviewers' general leniency toward lower-quality work.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Do LLMs Favor LLMs? Quantifying Interaction Effects in Peer Review
- LLM-Generated or Human-Written? Comparing Review and Non-Review Papers on ArXiv
- Mapping the Increasing Use of LLMs in Scientific Papers
- Has the Creativity of Large-Language Models peaked? —an analysis of inter- and intra-LLM variability —
- LLM-REVal: Can We Trust LLM Reviewers Yet?
- LLM or Human? Perceptions of Trust and Information Quality in Research Summaries
- Scientific production in the era of Large Language Models
- Stop Automating Peer Review Without Rigorous Evaluation