ABSTRACT
The Article’s starting point is a structural omission in the literature on AI responsibility. The three standard answers to the question of AI wrongdoing — the responsibility-gap literature (who is responsible when human control is too thin for blame to attach?), the legal-personhood debate (should an AI system be granted legal personality?), and the compensation-fund and insurance literature (how should victims be paid?) — disagree about the subject of AI responsibility but agree about its possibility. Each presupposes, without ever defending, that there is or can be a something on which a liability lands. This Article argues that this presupposition is the first question of AI law, and that it has been skipped: the first question is not whether artificial intelligence should be responsible, but whether it has any anchor at all.
The Article defines the missing object. A responsibility anchor is a bearer-position at which a sanction can land and which constitutes the subject’s own loss — so that imposing the sanction is an event for the subject, not merely an event in the world around it. The carrier principle states the discipline: responsibility is actual only where there is something that can be deprived; where there is nothing to deprive, responsibility is nominal — a declaration, not a liability. Three features follow. An anchor is not a thing (not the model, the weights, the chip, the data center — those are resources, and seizing a resource is confiscation, not sanction); it is a relation among a sanction, a continuity, and a stake, and a relation can be constructed by rule. Anchoring is a matter of degree, not a binary, so the law can build anywhere along a gradient. And the anchor is the material condition of answerability, not its substitute: attribution without anchoring produces a name with no body; anchoring without attribution produces a fund.
The Article specifies the anchor’s anatomy in four requirements — identity (the sanctioned entity is the acting entity, across time), stake, enforceability, and non-avoidability — and then undertakes the step it considers most consequential: an objective reconstruction of stake. A sanction satisfies the stake requirement not because the subject feels the loss, but because the subject’s acting depends on the path through which the sanction is imposed; the criterion is functional dependence, an auditable, indicator-keyed condition — not a verdict about consciousness. On this basis the Article distinguishes two tracks: the exogenous anchor (the regime’s registry, license, and account — buildable today, requiring no judgment about the system’s interiority) and the endogenous anchor (a system’s own stakes in its own continuation — conditional, future, and defined by the same auditable metric). It also answers the forking problem: because a model’s weights are a species reproducible at will, identity anchors to the registry record and the authorized operating unit — not to the weights — and the Article states the honest limit of that answer where the record is only observable and not yet attributable.
The conclusion is one sentence: give the machine a corpus, not a personhood. The Article makes four contributions. Conceptually, it identifies the responsibility anchor as the first question of AI liability, displacing the trio of gap, personhood, and payment, and names the failure mode — nominal responsibility. Doctrinally, it derives the four requirements and the seven selection tests as a unified criterion set, with stake reconstructed objectively. Historically, it reconstructs the three migrations and the two laws. Institutionally, it offers the access anchor and the allowance account as a design that solves answerability without personhood — buildable now, from existing law in four legislative steps, and structured so that if an AI with an inner life ever arrives, the law is already waiting with the forms.
He, Mingdong, The Responsibility Anchor – Why AI Law Needs a Bearer before it needs a Person (September 2, 2026).
Leave a Reply