Developing a Liability Framework for Harms Arising from Specification Gaming

This paper studies the development of a legal liability framework to address harms stemming from specification gaming in artificial intelligence (AI) systems. It argues for a two-step approach. Firstly, it examines the existing legal rules pertinent to commercialized AI products, particularly in contract, tort, and product liability, as well as the compliance standards concerning data and AI systems, which may serve as benchmarks for determining liability. Secondly, the paper proposes the formulation of new rules to tackle emerging new challenges posed by specification gaming, such as standards for effective reinforcement learning. Moreover, it suggests innovative compensation mechanisms, including the establishment of a dedicated fund to address incidents related to specification gaming.