• Sources: Unslop write-up, HN discussion
  • Summary: A write-up on 2026-07-20 described measuring the prevalence of AI-generated prose in arXiv papers and where the measurement method breaks down, covering marker-word frequency shifts, base-rate confounds, and false positives on non-native English writing.
  • Why it matters: Detection heuristics for machine-written text are increasingly load-bearing for review and moderation, and the post documents their failure modes.

send feedback on this story