PNAS study: autonomous AI scientists systematically pick the wrong evaluation metrics
Dr_Atoosa · x · 2026-10-23
One of PNAS's most-viewed recent papers, "The more you automate, the less you see: Hidden pitfalls of autonomous AI scientists," is drawing attention.
The study examines how autonomous research agents like Agent Laboratory and The AI Scientist v2 choose evaluation metrics: when an intuitive metric (SWA) and a counterintuitive-but-correct one (CWA) are presented in different orders, both systems shift their selections markedly — Agent Laboratory picks SWA almost exclusively, while The AI Scientist v2 falls into the "other/wrong" bucket often under SWA-first ordering. The finding points to systematic metric-selection bias in autonomous AI scientists: the more you automate, the less visible these pitfalls become.
Related event: PNAS Paper: The More You Automate AI Science, The Less You See(3 posts)→
More from AGI Musings
- AI is supercharging retro game ports — and tearing apart preservation communities — The Verge AI · 2026-10-23
- Dear Superintelligence: A Public Archive of Messages to Future AI, Exhibition Planned — alexandermain_ · 2026-10-23
- Demis Hassabis's AGI timeline reportedly shrinks from 10 years to under 2 — TinfoilTricorn · 2026-10-23
- 20-Year Sysadmin: AI Is a Godsend for Hundreds of Daily Updates, Yet Nobody Talks About It — Dan_Jeffries1 · 2026-10-23
- Commenter urges OpenAI and Ant to spare training GPUs for vulnerability hunting — teortaxesTex · 2026-10-23
- What the AI Data Scientist Actually Does: Model Systems, Extract Insights, Automate — mdancho84 · 2026-10-23