Inclusive AI in Administrative Services: Divergent Understandings and Governance for Inclusion

• In Event: AI, Equity, Ethics, and Inclusion

Abstract

As artificial intelligence is increasingly introduced into public services, questions of digital inclusion extend beyond access to technology to issues of governance and public value. In administrative service consultation and delivery, AI systems can facilitate access to information, provide guidance, and directly deliver services or make decisions affecting citizens. If inclusive AI is not properly designed, these systems may produce recommendations or decisions that fail to meet inclusion goals. Divergent understandings of what counts as inclusive AI among different actors can create misalignment between policy design and social expectations, undermining fairness and public acceptance, particularly for individuals with lower digital capacity or greater exposure to administrative exclusion.

This study addresses two related questions. First, do different actors understand inclusive AI in systematically different ways? Second, which governance arrangements can respond to these differences, foster a shared understanding of inclusive AI, and support the delivery of inclusive public services? Drawing on a public value perspective, the paper conceptualizes inclusive AI along three dimensions: accessibility, protection, and participation. Accessibility refers to whether diverse users can effectively access and use AI-enabled services. Protection refers to safeguards against unfair or exclusionary outcomes. Participation refers to opportunities for affected groups to voice concerns and influence how AI systems are governed. A benefit and risk distribution logic further explains why actors facing different expected benefits and risks prioritize these dimensions differently.

The study uses original survey data from China and Japan and employs a two-part design within a single survey. The first part develops and validates a measurement framework for inclusive AI and examines heterogeneity within the public, particularly by digital capacity, prior AI experience, and trust in government. The second part embeds a factorial vignette experiment with eight conditions, in which respondents are randomly assigned to scenarios of AI-embedded administrative service delivery. Scenarios vary by governance dimensions: who participates in rule setting, whether explicit protections for vulnerable groups are included, and whether the system is dynamically revised based on feedback. These manipulations aim to identify governance arrangements most effective in promoting inclusive AI, ensuring AI-supported services are equitable and responsive to all users.

Preliminary evidence from the pilot survey indicates that respondents distinguish among accessibility, protection, and participation when evaluating inclusive AI. Governance arrangements that include explicit protections for vulnerable groups and ongoing feedback-based adjustment are perceived as more effective in meeting diverse population needs than expert-led designs. Effects are particularly pronounced among respondents with lower digital capacity and lower trust in government, highlighting the role of actor heterogeneity in shaping perceptions and judgments of inclusive AI implementation.

This study demonstrates that inclusive AI should not be treated as a self-evident policy objective. Effective policy design requires recognizing that different actors prioritize different inclusion dimensions. By identifying governance arrangements perceived as legitimate and fair, the study provides practical guidance for policymakers designing AI-enabled administrative services that are efficient and responsive to vulnerable populations. These insights contribute to academic understanding of inclusive AI governance and to practical strategies for implementing socially responsible AI in public service delivery.

Author

• LIZHI YU, Tsinghua University
  Presenting Author