Human Perceptions on Moral Responsibility of AI: A Case Study in AI-Assisted Bail Decision-Making

Title of the Paper

Human Perceptions on Moral Responsibility of AI: A Case Study in AI-Assisted Bail Decision-Making

Paper Information

• Subject Area: AI Ethics, Moral Responsibility, AI and Decision Support
• Keywords: AI, Moral Responsibility, Accountability, Moral Judgment, Blame, Responsibility Attribution, COMPAS, Bail Decision-Making

Research Background and Problem

• Problem or Challenge: When AI systems cause harm, who should be held accountable for the consequences? The high autonomy and learning capabilities of AI complicate the issue of responsibility attribution, often referred to by researchers as the "responsibility gap." Current discussions on this issue largely focus on ethical principles and theoretical frameworks, but practical guidance remains scarce.
• Significance: The widespread use of AI has profound implications for law, ethics, and social order. Addressing the issue of responsibility attribution is not only critical for the ethical development of technology but also directly impacts human acceptance and trust in AI.
• Research Motivation and Related Work:
  • There is a growing body of opinion suggesting that designers should be held accountable for the behavior of AI systems, but debates persist on whether AI systems themselves can bear responsibility.
  • Many studies focus on philosophical and theoretical discussions of responsibility, with a lack of empirical research on public perceptions of moral responsibility attribution to AI.
  • This study aims to use survey data to gain deeper insights into how the public perceives the attribution of responsibility between AI and humans in high-stakes decision-making scenarios.

Solution

• Method or Solution:
  • Two experiments were designed to investigate public perceptions of responsibility attribution between AI and humans in bail decision-making scenarios:
    • Experiment 1: Examined moral responsibility attribution when AI or humans act as advisors.
    • Experiment 2: Explored responsibility attribution when AI or humans act as final decision-makers.
  • Collected public evaluations of eight responsibility concepts derived from philosophical and psychological literature, including "responsibility-as-task," "responsibility-as-authority," "responsibility-as-obligation," among others.
  • Used real-world data (e.g., bail recommendation data from the COMPAS tool) to construct experimental scenarios.
• Innovative Aspects:
  • Holistically measured public perceptions across multiple responsibility attributes, rather than focusing on a single domain (e.g., blame or punishment).
  • Combined real-world AI application contexts with experimental design to enhance the study's realism.
• Implementation Steps:
  • Employed factorial design methodology, testing responsibility attribution through randomly selected cases, and observed participants' quantitative evaluations of responsibility concepts under different conditions.
  • Analyzed data using linear mixed-effects models to control for individual differences among participants.

Research Findings

• Specific Findings:
  • The public believes both AI and humans should be held accountable for their actions, but the forms of accountability differ:
    • Humans are perceived as more responsible for "task completion" or "supervisory responsibilities," which are forward-looking responsibilities.
    • There is no significant difference between AI and humans in backward-looking responsibilities such as "causal responsibility," blame, or compensation.
  • Whether as advisors or decision-makers, the public expects both AI and humans to explain their decision-making processes, highlighting a widespread demand for explainable AI.
• Advantages:
  • Compared to existing theoretical research, the empirical data from this study provides valuable references for AI ethics and policy design.
  • Highlights public perceptions of AI's actual responsibilities in human oversight and legal domains, contributing to the optimization of human-AI collaboration models.
• Experimental or Evaluation Results:
  • Across both experiments, humans were generally perceived as having the ability and obligation to complete tasks, while AI was attributed these responsibilities to a lesser extent.
  • In areas such as blame and causal attribution, AI was perceived as lacking "human qualities," but its level of responsibility was nearly equivalent to that of humans.
• Limitations and Future Directions:
  • Limitations:
    • The experiments were based on specific scenarios (e.g., bail decision-making), and the results may not generalize to other domains such as healthcare or autonomous driving.
    • The study did not address situations where disagreements might arise between AI and humans (e.g., discrepancies between advice and execution).
  • Future Directions:
    • Explore responsibility attribution in broader scenarios, particularly in high-risk domains.
    • Investigate how public perceptions of responsibility can be embedded into AI design processes, such as the standardization of explainable algorithms.
    • Examine additional responsibility concepts, especially those not widely applicable to AI systems, such as virtue responsibility.

Conclusion

This study reveals the public's varying perceptions of responsibility attribution between AI and humans in high-stakes decision-making, emphasizing the need for explainability and moral responsibility in AI systems. It provides insights for the ethical design and policymaking of AI technologies.