Pangram scan: fewer than 5% of ICLR reviews look human-written
HildeKuehne · x · 2026-11-13
Hilde Kuehne shares her workflow for ICLR rebuttals: her student ran every question-answer pair through the Pangram LLM-detection tool to decide which reviews get answered in LLM style and which in human style. The result: no review scored even 5% likely-human, meaning virtually none of the reviews were written without AI. She asks human reviewers to identify themselves so authors can send short, real answers instead of gambling on style.
More from AGI Musings
- Kokotajlo tells middle powers to loudly treat superintelligence as real — emergency declarations, summits, audits — connoraxiotes · 2026-11-13
- Tech history comes in brief golden epochs: Athens, Renaissance, Bell Labs — wavefnx · 2026-11-13
- John Hempton: AI needs $2T yearly revenue, implying $4T in replaced wages — marigo · 2026-11-13
- Fable 5.2 'is AGI': a 20-part carousel makes the long case — BLCNYY · 2026-11-13
- 'India AI' Is Everywhere in Discourse, But Nobody Can Define It — vaibhavbetter · 2026-11-13
- Privacy scholar Carissa Veliz: Huang wrongly treats every problem as an engineering one — CarissaVeliz · 2026-11-13