Abstract
<title>Abstract</title> <p>Artificial intelligence deployed in public institutions increasingly aims to be socially intelligent: reading emotional cues, adapting to context, and responding with apparent empathy. Existing critical scholarship on algorithmic governance in welfare, education, and social services has focused largely on an earlier generation of opaque, extractive risk-scoring systems, exemplified by the Dutch SyRI algorithm and the Australian Robodebt scheme. That literature treats the citizen-state relationship as instrumental: a problem of transparency, bias, and due process. It has little to say about a newer design paradigm in which public AI is built to converse, to remember, to mirror affect, and to be experienced as a relational presence. Drawing on parasocial-interaction theory, media-equation research on social responses to machines, and recent empirical work on AI companionship, this paper proposes the construct of parasocial governance: the structural condition that arises when socially intelligent public AI cultivates relational trust and disclosure in citizens without a matching architecture of relational accountability, namely continuity, confidentiality, duty of care, and contestability. A typology is developed, distinguishing disclosed instrumentality, accountability-paired relational scaffolding, and parasocial exposure, and the mechanism is traced through documented cases spanning a state-funded elder-companion robot program, an NHS-commissioned mental-health chatbot, and litigated harms from commercial AI companions that are now shaping regulatory expectations. The paper closes with a research and policy agenda for accountable relationality in the design of public-facing AI, addressed to the welfare, health, and education systems in which socially intelligent interfaces are now being adopted.</p>