How we measured AI writing across arXiv, and where the measurement breaks

We scored the full text of 12,750 arXiv papers and found that about a third of new ones read as machine-written. Here is the method, the results, and an honest account of the limitations.