Human-AI Interaction Models for Socially Responsible Automation

Abstract

Socially responsible automation requires human-AI interaction (HAI) designs that maximise automation benefits -- efficiency, accuracy, scale -- while preserving the social, relational, and political dimensions of human decision-making that automation threatens to erode. The design of the interface between human judgment and AI capability -- the HAI model -- determines whether automation augments human agency or substitutes for it, whether it concentrates power in deploying organisations or distributes capability to users, and whether it maintains human skill and social competency or allows them to atrophy. This paper proposes the Human-AI Interaction for Responsible Automation Framework (HAIRAF), a systematic evaluation of six HAI models -- advisory AI, collaborative AI, supervisory control, human-in-the-loop, human-on-the-loop, and fully autonomous with audit -- across five social responsibility dimensions: human agency preservation, skill retention, power distribution, social relationship maintenance, and democratic legitimacy. HAIRAF evaluates each HAI model across four deployment contexts: knowledge work, care relationships, civic participation, and creative production. Key results: advisory AI achieves the highest social responsibility score (0.904) by preserving human agency while providing AI capability augmentation; human-in-the-loop achieves the best balance of efficiency and social responsibility (0.876); fully autonomous with audit achieves highest efficiency but lowest social responsibility (0.608) due to skill atrophy, power concentration, and democratic legitimacy concerns. The framework provides HAI model selection guidance for socially responsible automation deployment.