Abstract: Over the past decade, the burgeoning development of artificial intelligence technologies and applications has brought about revolutionary changes to human society. Algorithms have been widely adopted in public decision-making and services, becoming a significant variable within the social power structure. Algorithmic power represents an internalized, diffused relational dynamic that exacerbates the risk of public policy failures as it permeates this domain. Therefore, this thesis, through a review of literature, sums up the concerns of scholars both at home and abroad in their discussions about governmental algorithm applications. It aims to clarify the inherent logic behind the embedding of algorithmic power in public policy. Moreover, this research reflects on a series of risk manifestations and their impacts from the perspective of three core issues in public administration theory: misalignment of public values, multiple-agents issues within organizations, and problems in the exercise of administrative discretion. In light of these findings, the thesis advocates that governments construct an evidence-centered, empirically-driven decision-making mechanism and explore collaborative governance through an algorithmic regulatory sandbox model, and administrative personnel should enhance their capabilities in algorithm application.