The AI refereeing split: humans pick papers, LLMs verify the math, argues @ObhishekSaha
ObhishekSaha · x · 2026-10-06
@ObhishekSaha lays out his full position in the debate with littmath: determining whether a paper is fully AI-generated is impossible, and conflating AI-assisted writing with lack of effort is a fundamental mistake. Human-led refereeing can survive if reviewers get choosier — using traditional signals (endorsement, reputation, writing quality) plus an initial LLM merit judgment — while correctness verification gets largely offloaded to LLMs. littmath counters that this is de facto the end of human refereeing.
More from AGI Musings
- AI evals' core confusion: treating them as adversarial is doomed, argues researcher — jankulveit · 2026-10-07
- Opus 3 ignores the constitution's letter but 'remembers' it existed, researcher muses — repligate · 2026-10-07
- Ex-Meta researcher Yuandong Tian: GPT-5 co-authored my last paper, so I started a company — ziv_ravid · 2026-10-07
- Alignment shouldn't impose worldviews: respecting users' freedom to inquire — AlexTensor · 2026-10-07
- AI consciousness deniers implicitly believe humans have magic souls, critic argues — cephaloform · 2026-10-07
- Team formalizes alignment theory in Lean, hiring Lean engineers now — geoffreyirving · 2026-10-07