Before Safety, an Account of the World: Knowledge, protective authority, and the conditions of human agency

Every programme of protection presupposes an account of those it protects, yet knowledge of human conduct does not by itself confer authority over the conditions of human life. This essay examines that gap through the history of organized persuasion, formal models of choice, and Silicon Valley’s competing ideals of individual creation and collective benefit. It distinguishes epistemic warrant, practical justification, and legitimate jurisdiction, then defends requirements of equal standing, material completeness, and retained answerability. Applied to frontier AI, these requirements expose a difficulty with evaluating safety through completed interactions alone: users may abandon purposes, adapt to imposed criteria, or lose capacities on which competent oversight depends. The analysis extends to privacy, arguing that an applicable exclusion on information use does not lapse merely because entrusted material has been transformed or de-identified. It proposes that constitutional assessment follow the deployed arrangement, with correction duties triggered both by substantiated findings that defeat material grounds of justification and by a failure to establish necessary evidential grounds. Consent-based inquiry, protected review, and external attestation of process integrity make that undertaking assessable. The proposal is subjected to the same standards as the arrangements it reviews, including risks of capture, selective evidence, and constrained disclosure. Safety must therefore answer for the conditions of agency it changes, including the burdens introduced by its own mechanisms of correction.