AI Scientists Face Fundamental Hurdles for Autonomous Discovery

A new position paper from arXiv (2605.08956) argues that agentic AI scientists are not built for autonomous scientific discovery, despite functioning as co-scientists. The authors identify four key challenges: problem selection suffers from the McNamara fallacy; large language models (LLMs) lack tacit procedural and failure knowledge from laboratory practice; preference optimization during post-training reduces output diversity toward consensus; and scientific benchmarks focus on single-turn prediction without feedback from physical experiments. These issues require revisiting fundamental design choices, not just scaling or scaffolding.