Full-stack alignment: co-aligning AI and institutions with thick value models

Aligning only with operator intent cannot guarantee that AI brings about good social outcomes; the paper proposes "full-stack alignment": aligning AI together with the institutions that shape it to the values people care about. Value representations such as utility functions and preference rankings struggle to distinguish values from other signals, support normative reasoning, and model the collective good; the authors advocate "thick value models." Five demonstration directions: AI value guardianship, normatively capable agents, win-win negotiation systems, meaning-preserving economic mechanisms, and democratic oversight institutions.