LEGAL ACCOUNTABILITY OF AUTONOMOUS AI SYSTEMS IN CRIMINAL JUSTICE

Abstract [English]

The speed of change caused by autonomous AI systems poses a basic problem for principles of criminal law. When an AI system takes an action that is normally a crime if taken by a human being, it becomes far less clear as to how to impose liability under law. Current legal frameworks, which are based on concepts of human agency, intent, and responsibility, lack the legal suitability to apply these authorities to an act that has autonomously been executed, with particular attention to concepts of AI blackness and autonomy. Further muddling the issue is the way in which many modern AI systems are "black boxes," especially deep learning systems that learn and evolve in unintelligible ways even to their inventors. Given this black box characteristic, it becomes nearly impossible to track a harmful action to a particular programming mistake, dataset flaw, or decision. Without this ability to show that a human actor (developer, agency owner, or consumer) caused the harmful output of AI, responsibility will be diffused along the chain of actors as it will be exceedingly hard to match the harmful output to a single actor responsible for that harmful output. The arrival of autonomous artificial intelligence systems necessitates a reconsideration of basic legal concepts. Actus reus traditionally relies on a morally culpable human actor - and it falls short of addressing AI harm. While there are serious legal complications with blaming the programmer or user, corporate criminal responsibility for manufacturers is a more feasible option. As AI continues to evolve, our legal systems must adjust their approach to ensure accountability, protect the public from harm, and fairly allocate responsibility when the distinction between human and machine actions erodes.