人工智能产品算法设计者的犯罪过失判断——以危惧感说的核心观点为立场

判断人工智能产品算法设计者的过失时,修正的旧过失论对导致危害结果发生的因果历程设置了具体预见可能性标准,这一标准与算法的黑箱属性以相关性而不是因果关系为基础的决策过程存在冲突。同时,此立场只重视结果忽视行为的逻辑会打击算法设计者的积极性,阻碍算法进步。新过失论虽然以结果避免义务作为犯罪过失的核心,但是其对结果预见可能性的标准缺乏具体设计,在判断预见可能性时往往束手无策。因此,两种立场都不是判断人工智能产品算法设计者犯罪过失的合理方案。相比之下,虽然危惧感说(超新过失论)认为结果预见可能性只需达到危惧感的观点受到了主流观点的批评,但这种批评值得商榷:其一,只看到了这种立场对结果预见可能性的低程度要求,却没有看到这种要求背后的核心观点对于判断犯罪过失的合理性;其二,将危惧感说提出者本人对个别案件的过失判断等同于危惧感说的全部,略显片面。与修正的旧过失论和新过失论相比,危惧感说的核心观点是:结果预见可能性与结果避免义务存在相互关联性,这是判断人工智能产品算法设计者犯罪过失的合理方案。以危惧感说的核心观点为思路,犯罪过失包括客观的结果预见可能性、客观的结果预见义务和客观的结果避免义务。人工智能产品算法设计者客观的结果预见可能性的标准是:一旦遭遇包含异常因素的特殊情况,算法有可能做出不利决策,进而引发消极后果。算法设计者客观的结果预见义务的内容是:其一,应当预见到其设计的算法不仅会被用于没有异常因素出现的正常情况,而且可能被用于伴随异常因素出现的特殊情况;其二,一旦其设计的算法面临特殊情况,该系统可能会做出不利决策。算法设计者客观的结果避免义务的内容是:应当避免在设计算法时植入为社会公众普遍反对或不赞同的价值理念;在设计时检验"投喂"给算法系统的数据质量,最大程度防止缺陷数据进入算法机器学习训练的"垃圾进"风险;及时告知产品生产者算法可能面对的异常情况。

The criminal negligence of the designer of artificial intelligence products: Taking a stand on the central perspective of the sense of fearing theory

When judging the negligence of the designer of artificial intelligence product algorithm, the opinion of modified theory of old negligence is in conflict withthe nature of black box of algorithm that depends on correlation rather than causationwhen making decisions; the logic of this standpoint that only valuing results while ignoringconductmay blow algorithm designers’ enthusiasm, impeding algorithm’s progressing; Although the new negligence theory takes the obligation of result avoidance as the core standard of criminal negligence, it lacks specific design of the standard of possibility of foreseeing and is often at a loss when judging the possibility of foreseeing. Therefore, both standpoints are not reasonable schemes to judge the criminal negligence of intelligent products algorithm designers. In contrast, although the perspective of the sense of fearing theory (hyper new negligence theory) holds that the possibility of results foreseeing only requires the conductor to have a sense of fearing about the harmful result is enough, which is criticized by the mainstream view, this criticism is worth of discussion. Firstly, it only sees the fear requirement on the surface of this stance, but it does not see the core view behinds of the perspective of the theory of fearing: The correlation between the possibility of foreseeing and the obligation of avoiding results. Secondly, the opinion that the author’s judgment on individual cases is equivalent to the theory of fearing itself is an overgeneralization. Compared with the modified old negligence theory and the new negligence theory, the core point of the theory of fearing that there is correlation between the possibility of the foreseeing of result and the obligation of result avoidance is a reasonable scheme to judge the criminal negligence of the algorithm designer of intelligent products. Based on the core view of the theory of fearing, criminal negligence includes objective possibility of the possibility of foreseeing results, objective obligation of foreseeing results and objective obligation of avoiding results. The criterion for the objective the possibility of foreseeing results of the algorithm designer of artificial intelligence products is that the algorithm system is likely to make an adverse decision once it encounters a special situation containing abnormal factors, which may lead to negative consequences. The objective result avoiding obligation of the algorithm designer is that the algorithm system should be foreseen not only for the normal situation without abnormal factors, but also for the special situation accompanied by abnormal factors. Once the designed algorithm system encounters special situations, the system may make adverse decisions. The content of objective result avoiding obligation of algorithm designer is as follows: it is necessary to avoid implanting values that are generally opposed or disapproved by the public when designing algorithms; The quality of the data fed to the algorithm system should be checked at designing time to prevent the risk of "garbage in" of defective data into the algorithm machine learning training to the greatest extent, Informing the product producer promptly that the algorithm may face abnormal conditions.