ICML position track rebuttals largely AI-written, argues for track-wide score calibration
sethlazar · x · 2026-09-25
- Researcher Seth Lazar (quoting Margaret Mitchell on rebuttals easily taking 30 hours) says a "huge proportion" of rebuttals in the position paper track were 100% AI-generated.
- He argues reviewers of a given paper are poorly calibrated: borderline scores act as an attractor for hedging reviewers, so papers cluster at that level and borderline accepts should be calibrated against the whole track — likely meaning many rejections.
- Even for non-borderline papers, he says it's sometimes necessary to overrule credulous reviewers, including those who violated policy by using AI for their reports.
More from AGI Musings
- Google engineer quits over AI chip work, saying AI is moving too fast — basedjensen · 2026-09-25
- Amazon blocks Meta's shopping AI agent, plans to block Google's and OpenAI's too — SnoozeDoggyDog · 2026-09-25
- The decade-old worry that next-token representation would hinder learning turned out unfounded — Aaroth · 2026-09-25
- Why 'LLMs are just next-token predictors' is a contentless argument — Aaroth · 2026-09-25
- 'Next-token predictor' is a contentless way to describe LLMs — Aaroth · 2026-09-25
- Bengio on building AI: "What countered it is love, love of my children" — JacquesThibs · 2026-09-25