Product Safety in the Age of AI: Autonomy, R&D, and Liability

Abstract We study optimal liability for AI-powered products. Like human users, artificial intelligence (AI) can cause product failures that harm third parties. Additionally, it may introduce extreme risks of large-scale harm that renders full liability impractical. Raising AI liability for ordinary loss above actual harm can decrease excessive autonomy and increase social welfare, even when it negatively impacts R&D efforts. A well-designed liability rule implements efficient levels of autonomy and balanced R&D that reduces AI’s general risk. However, under targeted R&D to reduce AI’s extreme risk, full efficiency cannot be achieved with liability, and regulations limiting such risk can perform better.