Position: Falsify, Don’t Just Discover - AI Generated Discoveries are NOT Born Scientific

Zijun Liu, Kaiming Liu, Yiqi Zhu, Xuanyu Lei, Zonghan Yang, Zhenhe Zhang, Peng Li, Yang Liu

TL;DR: AI research systems need to incorporate explicit falsification process to generate scientific discoveries with effectiveness, responsibility, and ethics.

Abstract: Rapid development of artificial intelligence has drastically accelerated the development of scientific discovery. Recently, the rise of Large Language Models (LLMs) has led to the prosperity of autonomous agents, which enable scientists to seek references at different stages of their research. The demonstrated autonomy of these agents has led to designations such as "AI Scientist". However, it remains an open question whether we have truly reached the stage where scientific discovery can be fully automated. In this paper, $\textit{we posit that automated scientific discovery needs \textbf{automated falsification}, a missing part in the current research}$. As stated in Popper (1935), the central component of scientific research is falsification, where experiments are designed and executed to validate or refute hypotheses. To automate scientific discovery, the falsification process should also be automated. We review the substance of falsification in each stage along the development of AI-accelerated scientific discovery, and analyze the subject, the object, and the degree of automation of the falsification process. Following this, we initiate $\textbf{Baby-AIGS}$, a proof-of-concept AI-generated discovery system enabled by automated falsification. Through qualitative and quantitative studies, we reveal the feasibility of automated falsification, and advocate for responsible and ethical development of such systems for research automation.

Primary Area: Research Priorities, Methodology, and Evaluation

Keywords: AI-Powered Falsification, Automated Research, AI Researcher, AI Generated Discovery, Automated Machine Learning

Submission Number: 108