minntelligence
Research
Ongoing measurements of AI-generated content in public records, verified with Pangram 4.
National follow-up, in progress. We searched all 3,235 U.S. counties and county-equivalents for obtainable public-comment records; 1,883 publish them in machine-readable form. So far 3,565 comments across 221 local studies have been scored whole with Pangram 4: 5.7% flagged AI or Mixed (141 AI, 61 Mixed), the rest Human. Companion corpora scored the same way include federal SEC comments (1.1% flagged), ICLR peer reviews (0% before ChatGPT vs 12.5% after), arXiv papers (47.3%), Wikipedia drafts (31.3%), enacted laws and ordinances (0.3%), NGO/UN reports (5.6%), ChinaXiv preprints (88.7%), and Soviet-era papers from 1976–77 as a translation baseline (3.8%). Full study coming soon.
Pilot, complete. We scored the 20 public comments in the Metropolitan Council’s 2050 Transportation Policy Plan Amendment 1 report (comment period July 10–August 27, 2025). Of the 13 comments long enough to score, 1 was flagged as AI-generated at high confidence and 12 scored Human; 7 were under Pangram’s 50-word minimum. Two independent scoring runs agreed. Full study coming soon.
AI detection is probabilistic, not proof; a flag means the text statistically resembles AI output.