~25% of NBER working papers contain AI text; one program hits 32%, per Pangram analysis
soumitrashukla9 · x · 2026-09-10
Economist Paul Goldsmith-Pinkham used the commercial detector Pangram to measure AI writing in economics papers. Findings: almost no 2022 NBER or arXiv econ papers are classified as AI; today 50% of arXiv papers have at least one 300-word window flagged as AI and 25% of NBER working papers; the detected share takes off in 2025 with agentic AI, reaching a striking 32% in the Productivity, Innovation & Entrepreneurship program. The post also frames AI detection as a classification problem with false positive/negative trade-offs.
More from AGI Musings
- AI execs' common refrain: things are scary but we're trapped in a prisoner's dilemma — SatelliteNetSec · 2026-09-11
- Richard Socher: goal setting is the key to controlling superintelligent AI — RichardSocher · 2026-09-11
- Defending AI risk concerns: the marketing-conspiracy theories don't hold up — AndyMasley · 2026-09-11
- AI to cure all diseases in 5 years then end humanity, quip sparks mockery — MaxUnfried · 2026-09-11
- Is quantifying uncertainty arbitrary? A Bayesian ideology debate — NathanpmYoung · 2026-09-11
- Zvi disputes claim that AI lost a live debate against conservatives in Dallas — TheZvi · 2026-09-11