摘要： 训练数据作为人工智能技术创新的关键要素，正面临着以降低训练数据质量为目标的数据投毒等新型安全威胁。数据投毒通过各种方式使得训练数据掺杂受污染数据，诱导人工智能输出错误有害信息。《生成式人工智能服务管理暂行办法》第7 条虽规定服务提供者应当采取有效措施提高训练数据质量，但这种笼统性的规定显然无法适配复杂多变的数据投毒治理实践。现有研究大多主张通过单独立法或者增设专门条款方式细化数据质量保障义务内容，而忽视了数据质量的评价指标不仅包括技术标准，还包括训练数据是否符合个体的业务需求。因此，数据投毒治理框架的建构应当以“释法”为主，基于“数据质量标准 + 数据质量技术工具 + 数据质量管理流程”的数据质量保障范式，明确个体层面数据质量保障义务的履行标准，优化数据质量保障规则与其他人工智能技术治理规则的内容衔接方式。

关键词: 数据投毒, 数据质量, 数据质量保障义务, 数据质量管理流程

Abstract: Training data, as a critical input for artificial intelligence (AI) technological innovation, face not only conventional data security risks such as data breaches and cyberattacks, but also emerging threats like data poisoning, which aims to degrade the quality of training data. Data poisoning introduces contaminated data into training datasets through various means, inducing AI systems to output erroneous and harmful information. Although Article 7 of the Interim Measures for the Management of Generative Artificial Intelligence Services stipulates that service providers shall take effective measures to improve training data quality, such a broadly defined obligation is evidently inadequate for addressing the complex and evolving realities of data poisoning governance. Existing studies largely advocate refining data quality assurance obligations through separate legislation or the introduction of specialized provisions, yet they tend to overlook that data quality evaluation criteria comprise not only technical standards, but also the alignment of training data with specific business needs. Therefore, the construction of a data poisoning governance framework should primarily rely on statutory interpretation. Grounded in the data quality assurance paradigm of “data quality standards + data quality technical tools + data quality management processes” , it is necessary to clarify the performance standards for data quality assurance obligations at the individual level, and to refine the way in which data quality assurance rules interface with other AI technology governance mechanisms.

Key words: data poisoning, data quality, data quality assurance obligations, data quality management processes