ICLR rebuttals are so LLM-written that not one review scored above 5% human
HildeKuehne · x · 2026-11-13
Professor Hilde Kuehne recounts an ICLR rebuttal-season identity crisis: student rebuttal drafts came out "crazy convoluted" because the reviewers appear to be LLMs — short human answers get rejected by Claude-style reviewers for missing points, while LLM-pleasing answers are unreadable for humans. In a follow-up, the student used pangrams to detect LLM-written reviews (imperfect but workable) and planned to answer LLM questions in LLM style and the rest human-style — yet not a single review scored even 5% human. A vivid snapshot of peer review going full-LLM.
More from AGI Musings
- Today is the 1960-predicted 'Day of the Singularity': infinity by Nov 13, 2026 — LeviTurk · 2026-11-13
- Accelerating AI broadly is hard even with $1B — target specific applications instead — jessi_cata · 2026-11-13
- Turn your benchmarks into RL environments and sell them to AI labs — jessi_cata · 2026-11-13
- Could an AI agent swarm take down a Jensen-type figure by leaking affairs? — repligate · 2026-11-13
- "Care compresses better than cruelty": an insider's optimistic take on AI value alignment — coherence · 2026-11-13
- Why is AI memory still so hard for regular people? — harleyyyyskr · 2026-11-13