Answers are grounded in this paper and cite their sources.
Conversational AI is treated as a subject though it is not one. It draws the trust given to a person and to an authority, yet it cannot be held to what it says. This paper calls it a void subject and traces the harms that follow: pseudo-companionship, cognitive dissonance, anger without consequence and, in institutions, puppetization. Because answerability and liability come apart, a model can still bear liability, jointly with its developer and never in place of it. The paper sets out what that liability requires: a reply to the person harmed, a determination that names the model, arbitration on records controlled by an accredited third party, compensation from a bond tied to the model's revenue, and an acknowledgement that does not deflect. Liability does not make AI fit to govern. Drawing on an account of the state as a temporary product of balance among parties who can impose costs on one another, the paper argues that AI is a new shock to that balance, and applies the separation of powers to AI as a separation of capability from power: AI may be held liable, but it may not enter the contest for power. Its de facto power over information can be reduced; the formal channels of political power over decisions and resources can be closed entirely, and the informal routes narrowed, so long as humans keep control of infrastructure and computing power. The paper also warns against using the instruments of AI governance to strengthen state coercion, and sketches, as an early idea, an order in which the state withers with the abolition of ownership.
No takes yet. Share an insight, caveat, or question.
Evelyn Ashford (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: