作者：皮勇
作者介绍：同济大学法学院教授
出版日期：2026年4月9日
内容提要 对于人工智能系统自主决策和控制造成的危害后果，现行法律规制存有空白。若仅采取民事、行政法律手段，不足以遏制此类严重危害，需要引入刑法治理。关于刑法治理的路径，人工智能系统犯罪主体说难以实行，而妨害人工智能安全管理秩序犯罪立法的路径能够兼顾人工智能安全与发展，其中，采取违反人工智能安全管理义务犯罪的立法方案能够有效遏制前述危害。人工智能安全管理义务是为防控人工智能安全风险而设定的特殊、有限的规范责任。为有效惩治严重违反该义务的行为，有必要在刑法中增设违反人工智能安全管理义务罪。在罪行构造上，应依据义务犯、风险分配等刑法学理论，以有效降低人工智能安全风险为标准，合理规定该罪的人工智能安全管理刑法义务、主体和罪过等要件。
关键词 人工智能安全 风险 安全管理义务 义务犯 规范责任
Abstract: In cases where autonomous decision-making and control by AI systems lead to harmful consequences, there are gaps in the current laws. Civil and administrative legal measures are insufficient to curb such severe risks, necessitating criminal law governance. Regarding the approach of criminal law governance, legislation on crimes that disrupt the order of AI safety management can balance both AI safety and development. Among these, adopting a legislative plan for the crime that violates AI safety management obligations can curb the harms. AI safety management obligations are specialized and limited regulatory responsibilities designed to prevent and control AI-related risks. To effectively penalize serious violations of the obligations, it is necessary to establish a crime that violates AI safety management obligations. Based on criminal law theories such as duty-based crime and risk allocation, this crime should be reasonably defined in terms of AI safety management obligations, subjects, and culpability to effectively mitigate AI-related risks.