Att straffa det oförutsebara – En rättsdogmatisk analys av personlig skuld vid autonoma systemfel

Sammanfattning

The following thesis analyzes criminal liability in accidents caused by self-driving vehicles. These autonomous vehicles are governed by advanced artifi-cial algorithms that make autonomous decisions, which challenges the funda-mental requirements and principles of criminal law. Among other things, the maxim "nulla poena sine culpa" (no punishment without guilt) suffers when the decision-making process at the moment of the act lacks transparency and insight. In light of this, the thesis specifically analyzes the extent to which engineers or system developers, through negligence, can be held liable for causing the death of another pursuant to Chapter 3, Section 7 of the Swedish Penal Code (Brottsbalken). Through this, the thesis illustrates how the lack of transparency in AI systems, hereinafter referred to as the "black box pro-blem," complicates the assessment of criminal liability. Using a critical legal dogmatic method, it is analyzed whether current criminal law is sufficiently designed to handle technical innovation where decision-making processes remain hidden within complex systems. The elements of a crime in criminal law are constructed on the basis that the perpetrator is a natu-ral person with an individual consciousness to make decisions. The artificial intelligence in self-driving vehicles creates a theoretical artificial perpetrator who lacks human consciousness and whose decision chains remain conce-aled. A central conclusion is that the black box problem prevents the determination of personal guilt. According to the principle of conformity, criminal liability requires that the individual had the ability and opportunity to comply with the law. Since the engineers behind the systems lack insight into the system's de-cision chains, the prerequisites for establishing negligence through personal insight and blameworthiness are lacking. Furthermore, the standard of proof "beyond a reasonable doubt" is analyzed to illustrate the obstacles faced by the prosecutor due to the lack of insight into the systems. The thesis highlights reasoning from the Supreme Court regarding the requirement that alternative courses of events must be excluded for criminal liability, which the inherent lack of transparency in AI systems prevents. The result of this is a liability vacuum in Swedish criminal law where the legislation cannot punish any hu-man actor for the systems' shortcomings. Finally, social adequacy (social adekvans) is analyzed as a potential ground for exclusion of liability, in order to illustrate how criminal law should relate to AI innovation. Despite the social benefits of AI, the thesis dismisses the applicability of the doctrine of social adequacy. When opaque decision-making processes in self-driving cars replace and erase human control, the risk-taking in traffic cannot be seen within the framework of what is socially accepted or justifiable. Lastly, potential solutions for the legislator are presented to address the pot-ential liability vacuum in the legislation. The thesis particularly highlights the introduction of requirements for Explainable AI (XAI). XAI creates the con-ditions to transform hidden algorithms into clear decision chains, which pro-vides better opportunities to demand criminal liability and prevents artificial technology from eroding the purpose of criminal law.