[Purpose/significance] How to equip AI systems including AI agents with the knowledge and skills and make them behave in a more humanized way is the core proposition of AI governance. AI training, especially humanin-the-loop AI training, is a valuable solution to this problem. [Method/process] Firstly, based on the procedure anal? ysis, the basic process of human-in-the-loop AI training is proposed. Secondly, the interaction models of AI training are proposed based on the human-in-the-loop idea and AI training process. Moreover, by analyzing hiring announce? ments of AI trainers posted on Indeed, Linkedin, Zhaopin, Lagou and Baidu Baipin,the job responsibilities and require? ments of AI trainers are discussed. [Result/conclusion] AI training should not be limited to the training of knowledge and skills of AI, but should also focus on the training in humanistic and social dimensions. In addition, AI training can be divided into pre-training and retraining, which respectively focus on the training of cognitive ability and humaniza? tion of AI. The basic process and interaction models of human-in-the-loop AI training explored in this paper will pro? vide theoretical references for the development of AI training, especially AI retraining.