Percy Liang: experts rate 1/3 of AI-proposed research questions genuinely interesting
StanfordAILab · x · 2026-10-28
Percy Liang highlights the After42 project testing whether AI can propose research questions, not just solve them:
- Out of 135 final questions across mathematics, statistics, theoretical CS, and physics, 29 domain experts judged 45 genuinely interesting (1/3), with wide variation across subfields.
- Key method: develop questions by trying to solve them.
- AI judges' top picks aligned with expert consensus but failed to consistently capture individual experts' preferences.
- The team expects this to improve rapidly and says it raises hard questions about the desired human-AI relationship in academia and industry; experiment data is released.
Related event: After42: AI-Generated Research Questions Win Expert Approval(2 posts)→
More from AGI Musings
- Security researchers weigh whether AI-driven economic disruption will fuel global cybercrime — joshua_saxe · 2026-10-28
- ~20% of ICLR reviews may be AI-generated: Heng Li pushes back, expert says AI review can beat average humans — anshulkundaje · 2026-10-28
- Noah Smith: AI is the second-to-last important technological advance of the human age — abhiadesai · 2026-10-28
- When tasks can only be verified in the real world, RL rewards become reality itself — yacineMTB · 2026-10-28
- Susan Zhang mocks 'China wins AI via 996 engineers' take: smarter meatsacks won't decide the future — suchenzang · 2026-10-28
- Longevity KOL predicts cell resets will be as routine as teeth cleaning by 2035 — rand_longevity · 2026-10-28