Alignment as Institutional Design: From Behavioral Correction to Transaction Structure in Intelligent Systems
arXiv:2604.13079v1 Announce Type: cross
Abstract: Current AI alignment paradigms rely on behavioral correction: external supervisors (e.g., RLHF) observe outputs, judge against preferences, and adjust parameters. This paper argues that behavioral corr…