Why the Fine, AI? The Effect of Explanation Level on Citizens' Fairness Perception of AI-based Discretion in Public Administrations

Document Title

Why the Fine, AI? The Effect of Explanation Level on Citizens’ Fairness Perception of AI-based Discretion in Public Administrations

Document Information

• Topic Area: Application of Artificial Intelligence in Public Administration and Perception of Fairness
• Keywords: Algorithmic decision-making, administrative discretion, informational fairness, distributive fairness, explainable AI

Research Background and Issues

• Research Background: Automated decision-making systems, including fully AI-based decisions, are increasingly being introduced into public administration. These systems have the potential to improve efficiency, reduce costs, and optimize service quality. However, in scenarios requiring subjective discretion, automated decision-making faces significant challenges, particularly in maintaining fairness without undermining public trust.

• Issues and Challenges:

  • Citizens are skeptical about the fairness of AI-based discretionary decision-making.
  • The impact of the content of AI decision explanations on citizens’ sense of fairness and system acceptance remains unclear.
  • There is insufficient research on fully automated decision-making, especially in administrative discretion scenarios.

• Research Motivation:

  • The widespread impact of AI decisions on citizens' lives and the potential power imbalance they create.
  • Citizens’ perception of fairness influences their acceptance of AI technologies, which is crucial for the success of digital public services.
  • Current regulations (e.g., GDPR) mandate explanations for autonomous decisions but do not specify standards or characteristics for such explanations.

• Research Question:

  • How do different levels of AI decision explanations affect citizens’ perceptions of fairness (informational fairness and distributive fairness) and their acceptance of the system?

Solution

• Methods and Design:

  • The study designed a scenario-based survey: participants watched video clips demonstrating citizens updating expired ID cards via self-service terminals equipped with AI-based discretion.
  • Three levels of explanation were provided: no explanation (Baseline), factors explanation (Factors Explanation), and weighted factors explanation (Factors and Importance Explanation).
  • Quantitative data were analyzed using mixed linear models, supplemented by qualitative insights from open-ended questions.

• Innovations:

  • Proposed a new analytical framework combining explanation detail levels with citizens’ AI literacy.
  • Validated the impact of different explanation levels on fairness perception and technology acceptance through public surveys, refining research on citizen user experience.

• Implementation Steps and Techniques:

  • Developed a prototype self-service terminal running AI algorithms to support video scenarios.
  • Randomly assigned participants to watch video scenarios with three different explanation levels.
  • Designed and conducted a survey combining quantitative measurements and open feedback.
  • Extracted key findings through statistical analysis and thematic analysis.

Research Findings

• Key Findings:

  • More detailed explanations (e.g., showing decision factors and their weights) significantly improved citizens’ perceptions of informational fairness and distributive fairness.
  • Providing more explanation content did not significantly affect willingness to accept the technology.
  • Citizens with higher AI literacy were more willing to accept the AI system and provided more positive evaluations of decision fairness.

• Qualitative Analysis Findings:

  • Explanations were perceived as fairer than no explanations, but citizens expected not only information provision but also rationality, personalization, and clear decision grounds in explanations.
  • Citizens generally preferred retaining some level of human involvement, especially in cases of disputes or errors.
  • Strong recommendations were made to allow citizens to appeal decisions or implement mechanisms resembling human-like tolerance.

• Advantages and Value:

  • Contributes to designing more humane and fair AI decision-making systems suitable for public administrative services.
  • Provides empirical evidence from a citizen perspective on how to design AI explanation functions to enhance trust.

• Limitations:

  • The study only explored low-risk discretionary scenarios; applicability to high-risk scenarios remains to be verified.
  • The sample was limited to German residents, with limited representation of geographic and cultural diversity.

• Future Directions:

  • Extend research to high-discretion, high-risk public service scenarios (e.g., child welfare, immigration and nationality).
  • Investigate differences in acceptance of AI discretion and explanation mechanisms across regions and cultural contexts.
  • Explore the distribution of roles between humans and AI in hybrid discretion and potential fairness conflicts.

Summary

Through quantitative and qualitative analysis, this study confirmed the significant impact of explanation levels and AI literacy on citizens’ perceptions, offering design recommendations for optimizing AI applications in public administration. The findings have practical implications for policymakers and technology developers.