Detection Finds AI-Generated Text in 24% of NBER Economics Papers

Economist Paul Goldsmith-Pinkham used the commercial AI detector Pangram to analyze economics papers, finding that 24% of NBER working papers now contain at least one AI-generated passage, versus near-detectable levels in 2022, with a sharp rise starting in 2025. The findings, circulated and debated by multiple economists, have shifted discussion from whether AI should be used to norms and quality assessment in academic writing.

Confirmed

• The analysis was conducted by Paul Goldsmith-Pinkham using the commercial detector Pangram.
• Core data: 24% of NBER working papers show at least one AI-generated passage; almost none detected in 2022 NBER or arXiv papers, with a surge from 2025.
• The "Productivity, Innovation, and Entrepreneurship" program shows 32% AI-generated papers, with the analysis hinting this is largely due to one prolific author.
• Per Josh Gans's retelling, nearly half of arXiv economics papers show AI-writing traces.
• Jeremy Nguyen's figures match the original study.
• Jessica Hullman and Josh Gans voiced concerns that over-reliance on AI can make papers harder to read even when AI does what reviewers asked.

Unconfirmed

• The "nearly half of arXiv papers" figure appears only in Josh Gans's retelling; the exact scope awaits the original post.
• The attribution of the 32% concentration to one prolific author is an inference by the analyst, unnamed and unresponded.

Why it matters

• The study provides quantitative evidence that AI has penetrated economics writing at scale, confirming a trend scholars had sensed.
• It has triggered debate over "human writing" norms and academic integrity.
• Josh Gans highlights shifting attitudes: more economists now find AI writing acceptable.
• Jeremy Nguyen argues that if AI-assisted research yields large gains, "AI-generated text = low-quality research" should no longer be the default assumption.
• Josh Gans's own year-long experiment since late 2024 adds an author's perspective: AI produces papers fast but tempts authors to pad weak ideas—papers with rigor but no insight that top journals reject.
• The 32% concentration in one program raises questions of authorship and accountability.

2026-09-10 ~ 2026-09-12 · 8 related posts

• ~25% of NBER working papers contain AI text; one program hits 32%, per Pangram analysis — soumitrashukla9 · 2026-09-10
• 32% of NBER working papers in one program are AI-generated — paulnovosad · 2026-09-11
• 24% of NBER economics papers contain AI-generated text, analysis finds — JeremyNguyenPhD · 2026-09-11
• Josh Gans: economics profession clearly finding AI writing more acceptable — joshgans · 2026-09-11
• [source] Economist's year-long AI paper experiment: rigorous theory, human insight — joshgans · 2026-09-11
• AI writing surges in economics papers, scholars warn it makes papers worse — sethlazar · 2026-09-12

2 near-duplicate retellings: JeremyNguyenPhD · joshgans