关键词: 科学智能（第五范式）, 人工智能科学家, 科学认识论, 科学自动化

Abstract: [Purpose/Significance] This study aims to explore the challenges that artificially intelligent (AI) scientists may bring to scientific epistemology. [Method/Process] Scientific discovery has long been of interest to AI researchers. The next big step in AI is the development of AI scientists. AI scientists should be able to independently motivate, make, understand, and communicate discoveries. Although the current robot scientists are still just a form of AI-driven automated experimental apparatus, and the best AI systems today cannot define their own hypothesis space and experimental design. At best, they can be considered to be a primitive form of AI scientists. Clearly, the specific path of AI-driven scientific research or the transition to AI scientists will inevitably be influenced by the frontier development of AI. Current AI systems must overcome the following major technical challenges: 1) making strategic choices in their research goals; 2) developing the ability to generate exciting and novel hypotheses in areas that push boundaries; 3) designing innovative approaches and experiments to test hypotheses that go beyond the use of prototype experiments; 4) focusing on and describing important discoveries in a way that can be understood by human scientists. The highly autonomous AI scientists can either make discoveries on their own or collaborate with other human and machine scientists to make Nobel-level discoveries. After reviewing the relevant AI applications in scientific research, this study illustrates the main characteristics of AI scientists and the two disruptive changes they bring about at the epistemological level: a leap in AI capabilities and AI for Science as the 5th paradigm of scientific research. [Results/Conclusions] The implications of AI for Science are revolutionary, but recent AI-driven explorations in scientific research increasingly support the possibility of its realization. In this situation, discussions on the epistemological issues of relevant sciences need to go beyond general philosophical debates and instead explore epistemological strategies for the coming scientific revolution in AI. In view of the coming scientific revolution in AI, this study proposes four strategies. First, we should pay more attention to the problems and solutions in the process of developing AI scientists. Second, the key to advancing the scientific revolution in AI is to find ways to eliminate factors that may lead to failure. Then, we use different strategies to achieve the scientific revolution of AI. Finally, we take advantage of metaphorical methods to help us develop AI scientists.

Key words: AI for Science, AI scientists, scientific epistemology, automated science