Abstract: This study aims to analyze the global public’s perception of AI risks, examining the types and spatio-temporal evolution of these perceptions. It also seeks to identify the factors influencing these perceptions, providing empirical evidence for the development of trustworthy AI globally. By conducting semantic analysis of 4.3 million posts on social media and categorizing temporal features, a global database is constructed to quantify AI risk types across different regions. The study employs convex hull analysis, geographic information systems (GIS), and integrated machine learning methods to reveal the regional heterogeneity, influencing factors, and spatio-temporal evolution of global public risk perceptions. The findings indicate significant regional heterogeneity in global public perceptions of AI risks, shaped by the diffusion effects of innovative technologies, resulting in diverse perception categories. Individual characteristics, political trust, and government effectiveness significantly shape public risk perceptions. These perception categories exhibit a covariant logic, transitioning from a singular to a pluralistic pattern.