Judicial artificial intelligence is divided into conventional artificial intelligence and professional artificial intelligence, the former is to directly transplant artificial intelligence that has developed and matured in the general field to the judicial field without special algorithm updates, the main purpose is to liberate judges from heavy "transactional" work, so they cannot intervene in the core content of trials; The latter is an artificial intelligence specially developed for the judicial field, such as the sentencing assistance system, which intervenes in the substance of the trial, which is the core of judicial artificial intelligence. The current practice status of judicial artificial intelligence is that conventional artificial intelligence is effective because of its solid foundation, but the development and use of extremely important professional artificial intelligence is not ideal, mainly because the research and development of professional artificial intelligence lacks the theoretical supply of legal research, which is concretely manifested as "too abstract but not enough concrete". The deep reason is that there is no deep integration of legal expertise and artificial intelligence technology knowledge, that is, "those who understand technology do not understand law, and those who understand law do not understand technology". At the macro level, under the space-time background of the unpredictable explosion of intelligence, the life 2.0 stage (cultural stage) or the era of weak artificial intelligence is still the stage in the present and foreseeable future for a long time, and the sentencing artificial intelligence pretending to be a "tool" should still be positioned to assist sentencing rather than determine sentencing, and based on the value connotation of the standardized reform of sentencing. It should be further oriented to normative aid rather than reference aid, the difference between the two is the binding force of the phased sentencing conclusion given by the intelligent system on the judge. At the micro level, the algorithm construction of intelligent sentencing system should be based on the principle of sentencing logic-led algorithm logic, and take phased sentencing with "common" attributes such as sentencing benchmark and illegal punishment as the field of action rather than the final sentencing conclusion (declaration of punishment) outside its ability. In addition, in order to prevent algorithmic black boxes, algorithmic discrimination and correlation replacing causation, it is necessary to make the algorithm of sentencing artificial intelligence open and the interpretation of phased sentencing conclusions.