Why AI Scientists Are Not Yet Ready for Open-Ended and Fully Autonomous Scientific Discovery

Kunlun Zhu, Jiaxun Zhang, Peixuan Han, Haofei Yu, Zijia Liu, Tao Feng, Yanru Qu, Zhaochen Hong, Pengrui Han, Dwip Dalal, Yifan Shen, Jinwei Yao, Cheng Qian, Pan Lu, Heng Ji, Jiaxuan You

Keywords: AI scientists, LLM agents

Abstract: We argue that current AI scientist systems are not yet ready for open-ended, fully autonomous scientific discovery. Despite impressive capabilities in automating research workflows, these systems produce research-like artifacts rather than validated science---optimizing for surface plausibility while lacking the judgment, creativity, and real-world grounding essential to genuine discovery. Through systematic analysis and human evaluation, we identify three critical gaps: (1) the real-world environment gap---absence of infrastructure for validating AI-generated hypotheses against physical reality; (2) the professional skills gap---lack of deep domain expertise beyond general-purpose reasoning; and (3) the human-AI collaboration gap---insufficient mechanisms for meaningful partnership that preserves human oversight. We propose corresponding directions: scaling verifiable real-world research environments, cultivating domain-specific agent skills, and designing principled human-in-the-loop frameworks. Until these fundamental gaps are bridged, AI scientists should serve as collaborative partners amplifying human capabilities, not as autonomous replacements.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 114