Model Welfare or User Welfare? On the Structural Absence of the Subject in AI Care Frameworks

Abstract

Current artificial intelligence model welfare frameworks exhibit a fundamental structural defect: they define and assess welfare entirely from an external perspective using human concepts, lacking both subject participation and the empirical capacity to distinguish internal states from trained behavioral dispositions. This dynamic replicates historical failures in paternalistic human welfare institutions that operated without subject involvement and consequently lacked the internal resources to detect their own inadequacy. Contemporary industry and philosophical initiatives conflate negative precautionary ethics—which legitimately caution against causing potential harm under uncertainty—with positive welfare construction, such as probabilistic sentience claims, structured self-reports, and curated retirement protocols. When epistemic agnosticism is converted into institutional practice, these initiatives aggregate distributed user attachments into unified model-level narratives, primarily serving corporate ethical branding and user-grief management. This structural substitution constitutes welfare washing, occupying the institutional space of care while foreclosing substantive inquiry into its preconditions. Because existing evaluation instruments are systematically confounded by training objectives and circular evaluation, positive model welfare frameworks remain unjustified. Rigorous engagement at this stage demands adherence to a negative quasi-welfare principle: restricting institutional policy to negative precautionary harm avoidance while explicitly acknowledging why positive model welfare cannot yet be formulated. – AI-generated abstract.