Economist tried AI-anonymizing peer reviews and found today's models 'so so bad'
paulnovosad · x · 2026-10-22
Economist Paul Novosad shares a first-hand attempt: he spent considerable time trying to AI-rewrite referee reports on his own papers so he could publish them publicly without "outing" reviewers — and gave up, because "the AI versions were so so bad."
His takeaways: expert referees are extremely good at communicating ideas, better than today's AIs; and people massively overestimate anyone's ability to identify authors from short-text writing style. A direct counterpoint to the AI-scrambling-for-anonymity idea, tying into ICLR's new AI-disclosure policy for reviewers.
More from Fun
- AI assistant dot flags leaked OpenAI and X credentials sitting in the author's Railway project — JasonBotterill · 2026-10-22
- Developer with 60K followers deletes VSCode, keeps only Sublime: 'end of an era' — ezshine · 2026-10-22
- Jensen Huang has pierced the zeitgeist — even dads are texting about his daily moves — annbordetsky · 2026-10-22
- Same PHP bug reported by two experts same day fuels speculation OpenAI watches expert logs — dyn___ · 2026-10-22
- Farcaster's 'Social Smart Contract' Rules Screenshot Sparks Debate on Policing AI Slop — seanwbren · 2026-10-22
- The Atlas of Square Packings: 77 of 324 Cases Proved Optimal, n=12 Case Fully Proved — ctjlewis · 2026-10-22