Researchers analyzed the full text of 12,750 arXiv papers and found that about a third of new publications read as machine-written. The study outlines the methodology behind these findings and provides a transparent look at the limitations where AI measurement tools currently break down.