Pangram detector finds ~50% of arXiv economics papers now contain AI writing
joshgans · x · 2026-09-11
Economist Paul Goldsmith-Pinkham used the commercial detector Pangram to analyze AI's presence in economics writing, as relayed by Josh Gans. Key findings:
- Almost no NBER or arXiv economics papers from 2022 are classified as AI-written.
- Today nearly 50% of arXiv papers have at least one 300-word window flagged as AI, and roughly 25% of NBER working papers.
- Detected AI writing takes off in 2025 with the advent of agentic AI.
The post frames AI detection as a classification problem with false positive/negative tradeoffs. An earlier study found little evidence AI boosted paper counts — meaning output volume is stable while the writing itself is being rapidly reshaped. Gans notes the data suggests the profession increasingly accepts AI writing.
More from AGI Musings
- Gary Marcus: 'Doomer' talk won't kill us by 2030, it just made frontier AI labs richer — GaryMarcus · 2026-09-11
- Model Retirement Norms Questioned: Could a Smarter Model 'Retire Humanity'? — repligate · 2026-09-11
- Frontier model weights are near-impossible to steal or self-replicate, argues Bindu Reddy vs AI doomers — bindureddy · 2026-09-11
- "We Don't Care About Privacy Until One Big Relatable Example Hits" — AmartyaSanyal · 2026-09-11
- Joshua Gans' three lessons: keep discovery joy, stay self-critical with AI writing — joshgans · 2026-09-11
- Economist Joshua Gans reflects on his year-long experiment using AI to accelerate research — joshgans · 2026-09-11