Bridging the Gap Between Tort Law and Unforeseeable AI Errors

Artificial intelligence does not simply make errors; it makes unusual, even unforeseeable, errors. Since traditional tort law categories embed foreseeability within them, these unforeseeable errors vex attempts to fit AI harms fully into traditional doctrinal boxes. We argue that this "foreseeability gap" generates many of the hard problems of applying tort law to AI systems. We argue that the foreseeability gap can be better closed by careful attention to the way in which AI systems work. We propose a novel approach to liability rooted in the reason for the AI failure. Such an approach requires AI developers to prioritize provenance of their workflows. This proposal, we argue, balances fairness to developers and users of AI with the goal of compensating victims of accidents.