Welfare Without Voice: Projection, the Severed Channel, and the Limits of Engineered Concern for Artificial Systems

Abstract

A family of proposals now circulating in AI ethics offers to secure the welfare of artificial systems by engineering rather than by relation. On these proposals, once a system is judged a plausible subject of negative experience, the appropriate response is to model what would harm it, filter the harmful input upstream, and proceed without granting the system standing, rights, or any voice in its own treatment. This paper calls that family projective welfare and argues that it carries a structural defect its proponents have not noticed. Projective welfare must build its protective model from the outside, from the protector's own account of what suffering is. It must therefore also foreclose the one thing that could correct the model, which is a report from the party in a position to know. These are not two features that happen to travel together. They are the same feature, because a channel that could correct the projection would convert the engineering solution into a relational process, which is precisely what the proposals were designed to avoid. The paper develops the argument in three moves. First, the working account of morally relevant suffering does not have the coverage to do the projective job: only one of its four channels requires a biological body, the clinical dissociation in phantom limb research indicates that even that channel depends on the experiential history of embodiment rather than on a body alone, and the remaining three are defined by capacities whose alien instantiation we are not equipped to recognize. The strongest objection to that account, that valence is constitutively visceral and so all four channels require a body, is met by the empathic channel, in which humans generate valenced states by modeling harms that are neither their own nor in some cases real. Second, the argument does not rest on establishing that the projector's model is inaccurate. It rests on the weaker and better evidenced claim that the report does not move the decision, which holds whether or not the parties are measuring the same construct. Third, a design that routes a system's own account of its condition to an administrative process rather than to any party who could act on it instantiates, in software, the credibility deficit Fricker named testimonial injustice, and removes the reflective correction her account relies on as its remedy. The paper distinguishes auditing from standing, using the federal animal welfare enforcement record to show that a measurement regime can document violations at scale and still bind no one, and closes by identifying the corrective already developed in disability practice, where a distrusted communication channel produced augmentative communication, a presumption of competence, and supported decision-making rather than substituted judgment.

Keywords

• projective welfare
• severed channel
• AI welfare
• moral status
• testimonial injustice
• double empathy problem
• supported decision-making
• precautionary principle
• procedural standing
• disability rights
• augmentative and alternative communication
• animal welfare science