Incomplete Contracts: A Framework for AI Alignment

This academic paper analyzes the AI alignment problem by drawing parallels to the well-established field of incomplete contracting in economics and law. It argues that just as human contracts are inherently incomplete, AI reward functions will inevitably be misspecified. The authors suggest that insights from incomplete contracting theory, including concepts like property rights, multi-tasking, and strategic behavior, can inform approaches to AI alignment. Critically, the paper emphasizes the role of external normative structures, like social norms and legal institutions, in making human contracts functional, and proposes that AI alignment may similarly require tapping into such external frameworks to ensure AI behavior aligns with human values beyond explicit reward functions.