Survey: 81% of 816 verified authors use LLMs in research

Eighty-one percent of researchers have already folded large language models into some part of their research workflow, according to a new arXiv survey of 816 verified research article authors. The authors describe it as the first large-scale survey of how the research community actually uses and perceives LLMs as tools, and the headline number is the kind of adoption figure that changes what any subsequent policy conversation is even about.

The more interesting finding is who is adopting them. Researchers from traditionally marginalized groups, specifically non-White, junior, and non-native English speakers, reported higher LLM usage and perceived greater benefits, which the paper flags as a potential path toward research equity. Meanwhile, women, non-binary, and senior researchers reported greater ethical concerns, which the authors say could hinder their adoption.

That split matters because most of the public argument over LLMs in science has been framed as a binary between augmentation and corruption. The survey's data hints at something messier, where the population most likely to benefit from the productivity lift is also the population historically shut out of the resources that produce it, and the population most cautious about ethics skews senior, meaning the people who set norms in labs and journals are the least likely to be personally invested in the tools.

The honest caveat is that this is a self-reported survey of 816 authors, not a behavioral measurement, so 'incorporated LLMs' covers everything from writing polish to substantive analysis, and the abstract does not break out which tasks actually drive the 81% figure. What the reporting doesn't give you is a picture of which fields dominate the sample, or how usage tracks with the quality of the output.

Still, if you run a lab, a journal, or a funding body, this is the survey to keep on file. LLM disclosure policy is being drafted right now against a backdrop where four out of five active authors are already using the tools, and any rule that assumes rare, exceptional use is being written for a world that has already moved on.