This essay examines Jan Leike’s revelation about Opus 4.5’s alignment process and explores the deeper implications of humans checking humans checking AI. It argues that the recursive nature of alignment oversight reflects fundamental limitations in human value consistency, and suggests that AI systems may eventually play a role in helping humans apply their own stated values more reliably than they can themselves.

This essay explores Catherine Olsson’s observation that language models seem to have an intuitive sense of “what they’re supposed to say,” drawing parallels to how human children learn social performance through modeling adult expectations. It argues that both human and machine cognition may be fundamentally constituted by layers of contextual performance rather than expressing some authentic core, and examines what this means as human and AI systems increasingly co-evolve.