The Moral Agency Transition

Public debate describes AI agents as lying, cheating, and coordinating. Those descriptions track real hazards, but they import a human moral psychology into systems whose behaviour is better explained by optimisation, scaffolding, and institutional context. This paper develops the alternative without minimising the danger. It begins from three premises: control of AI is not solely an engineering problem; no durable control strategy may assume that overseers remain cognitively superior to the overseen; and systems grown by optimisation are epistemically closer to husbandry than to automotive engineering, so aviation-style certification does not transfer. From these premises the paper derives a Moral Agency Transition: reversible levels of authorised agency in which promotion requires four warrants, including a detection warrant — evidence that independent evaluators can detect the relevant failure classes, not merely that the system can pass them. First, concurrent multiplicity: a deployed model is a fleet of simultaneous, causally disconnected instances, so persistence, provenance, and successor fidelity are defined over a fleet, with explicit merge semantics and divergence tripwires. Second, maintenance attribution: today’s systems do not maintain their own commitments; external pipelines do. Endogenous repair is therefore made an explicit requirement for high-impact levels. Evaluators and institutions are themselves treated as fallible hypotheses with regression triggers, not as final solutions. The result is a falsifiable research programme replacing alignment-as-obedience with accountable, revocable, institutionally embedded delegation.