AI Ethics in Public Policy

Abstract

Public sector adoption of AI systems for policy decisions -- welfare eligibility, criminal sentencing, child protective services, immigration processing, and public health resource allocation -- raises distinctive ethical challenges that differ fundamentally from private sector AI deployment because government decisions carry the coercive authority of the state and affect citizens who cannot opt out. This study presents a systematic analysis of AI deployment in public policy across 42 government AI projects in 18 countries, evaluating each on six ethical dimensions: transparency (public disclosure of AI use and logic), accountability (clear responsibility assignment and appeals mechanisms), proportionality (matching AI sophistication to decision stakes), equity (impact on vulnerable populations), democratic legitimacy (public participation in deployment decisions), and remediation (mechanisms for correcting AI-caused harm). Projects were scored 1-5 by three independent evaluators with expertise in technology policy (ICC = 0.78). The mean ethical compliance score was 2.48/5-- below the threshold of 3.0 considered operationally adequate -- indicating systematic governance deficits. Transparency was the weakest dimension: 64.3% of projects provided no public disclosure that AI was being used in the decision process. Accountability showed the widest variation (SD = 1.52): some projects had comprehensive appeals mechanisms while others offered no recourse to affected citizens. Equity analysis revealed that 38.1% of projects had conducted no disparity assessment prior to deployment. Projects in countries with existing AI governance frameworks scored 42.8% higher on composite ethics than those without (3.12 vs. 2.18; p < 0.001). A practical public sector AI ethics checklist derived from the highest-scoring projects and aligned with the OECD AI Principles is proposed.