摘要
人工智能技术发展的新趋势体现为对规模庞大且高质量训练数据的需求和依赖,但同时也给数据投毒的广泛入侵提供了可乘之机。数据投毒对象的特殊性、机器学习的自主决策性以及参与主体的多元性构成了罪名适用和刑法归责难题。在罪名适用时,应将不法行为统合在既有刑法体系之中,无须动辄增设新罪。尽管涉及数据犯罪的样态具有复杂性,但可以根据损害后果的性质直接化归为侵害传统法益的犯罪。在具体适用过程中,针对投毒行为侵害的两类对象,训练集数据和模型数据分别展开讨论。刑事归责时应考虑目的性与正当性的双重实现。在数据提供者、训练者、服务提供者以及使用者之间展开公平、有效的风险分配,考虑主体间的利益平衡,在可能和可期待的范畴内划定注意义务,以确保在实现刑法有效规制和预防目标的同时不会阻碍新技术的发展。而对投毒者而言,由于投毒者具有高于一般人的“特别认识”,故应对其持从严态度,践行主观归责理念,明确机器学习有限的自主性并不会阻碍归责。并且,即便机器学习的介入导致投毒者对犯罪因果流程支配力有所减弱,但只要事后能够证明结果与投毒行为具有条件关系,就能确认结果归属。
The new trend of the development of artificial intelligence technology is reflected in the demand and reliance on large-scale and high-quality training data,but it also provides an opportunity for the widespread invasion of“data poisoning”.The particularity of the objects of data poisoning,the autonomous decision-making of machine learning,and the diversity of participating subjects have formed the difficulties of applying charges and criminal law imputation.When applying charges,the illegal behavior should be integrated into the existing criminal law system,and there is no need to add new crimes at will.Although the forms of data crimes are complex,they can be directly attributed to crimes that infringe on traditional legal interests according to the nature of the damage consequences.In the specific application process,the two types of objects infringed upon by poisoning behavior,training set data and model data,are discussed separately.The dual realization of purposefulness and legitimacy should be considered in criminal imputation.Fair and effective risk allocation should be carried out among data providers,trainers,service providers and users,the balance of interests should be considered among subjects,and the duty of care should be defined within the possible and expected scope to ensure that the development of new technologies is not hindered while achieving the effective regulation and prevention goals of criminal law.As for the poisoners,due to their higher level of“special knowledge”than the general public,a strict attitude should be taken towards them,and the concept of subjective imputation should be practiced.It should be clarified that the limited autonomy of machine learning will not hinder imputation.Moreover,even if the intervention of machine learning leads to a weakening of the poisoner′s control over the causal process of the crime,as long as it can be proved afterwards that the result has a conditional relationship with the poisoning behavior,the attribution of the result can be confirmed.
出处
《天府新论》
2025年第4期38-48,152,共12页
New Horizons from Tianfu
基金
国家社会科学基金重点项目“完善党和国家监督体系研究”(编号:21AZD088)。
关键词
大模型幻觉
机器学习
财产犯罪
注意义务
主客观相统一
large model hallucination
machine learning
property crime
duty of care
unity of subjec⁃tivity and objectivity
作者简介
徐蕴杰,南京大学法学院博士研究生,研究方向:刑法学。江苏南京210093。
