Exploring the Association between Moral Foundations and Judgements of AI Behaviour

Title of the Paper

Exploring the Relationship Between Moral Foundations and Judgments of Artificial Intelligence Behavior

Paper Information

• Subject Area: AI Ethics, Moral Psychology
• Keywords: Moral Foundations Theory, Responsible AI, Automated Decision-Making, Ethical Principles, Guidelines

Research Background and Problem Statement

• Identified Problems or Challenges:

  1. How are the moral behaviors of Artificial Intelligence (AI) perceived and judged by individuals, especially in morally tense situations, and what disagreements exist among people?
  2. Current normative approaches in AI ethical guidelines fail to adequately consider the diversity in individual interpretations of moral behavior.

• Significance of the Research:

  • As AI applications become increasingly widespread, their behaviors are often imbued with moral significance. It is essential to systematically study how individuals make moral judgments and how their moral composition influences these judgments.
  • Investigating the predictive power of Moral Foundations Theory (MFT) can provide insights for designing AI ethical principles that better align with human needs.

• Motivation and Related Work:

  • Moral Foundations Theory has been successfully used to predict people's reactions to specific human behaviors (e.g., euthanasia, abortion), but it has yet to explain individual differences in judgments of AI behavior.
  • Current normative AI ethical guidelines (e.g., the Montreal Declaration for Responsible AI) have been criticized for failing to capture the diversity of human responses.
  • This paper aims to supplement this research field with an empirical approach, examining the practical impact of moral foundations on AI ethical judgments.

Proposed Solution

• Research Methodology:

  • Use the Moral Foundations Questionnaire (MFQ) to measure the individual moral sensitivities of 240 participants.
  • Construct six scenarios involving AI moral behavior and have participants judge these scenarios while identifying the most relevant moral foundations.
  • Analyze the data using Bayesian modeling and reflexive thematic analysis to examine how individual moral foundations influence their moral judgments of AI.

• Innovative Contributions:

  • First empirical validation of MFT's predictive capability in the context of AI behavior judgments.
  • Integration of quantitative and qualitative methods to explore how non-moral factors, such as technological cognition and anthropomorphism tendencies, influence judgments of AI behavior.

• Implementation Steps and Key Techniques:

  • Step 1: Randomly assign participants to complete two components—scenario judgments and the MFQ questionnaire.
  • Step 2: Use Bayesian modeling to analyze the relationships between key variables.
  • Step 3: Conduct qualitative analysis to extract latent thematic connections between participants' technological awareness and ethical judgments, forming an explanation of their behavioral logic.

Research Findings

• Specific Findings:

  1. Participants' judgments of AI behavior exhibited weak consistency, with significant disagreements in scenarios involving "fairness" and "authority."
  2. The "Care" moral foundation score of individuals was the strongest predictor of their judgments of AI behavior, while other moral foundations (e.g., "Loyalty," "Sanctity") contributed less to predictive power.
  3. Qualitative analysis revealed that individuals' understanding of technology is a critical non-moral factor influencing their ethical judgments. Participants with limited technological understanding were more likely to anthropomorphize AI and attribute responsibility to the AI itself.

• Comparison with Existing Solutions and Advantages:

  • Current AI ethical guidelines primarily adopt normative approaches and overlook individual differences. This study emphasizes the descriptive diversity and predictive capabilities of MFT.
  • The research highlights the connection between technological cognition and moral perception, offering a new dimension for AI system design and user education.

• Experimental or Evaluation Results:

  • Across all six scenarios, the average consistency of participants' moral foundation relevance ratings (Kendall’s W) was 0.28, indicating "limited consistency."
  • Bayesian modeling showed that individuals' sensitivity to the "Care" moral foundation was the primary variable predicting AI behavior judgments (effect size: 0.24).

• Limitations and Future Directions:

  1. Limitations:

     • Single cultural context: Participants were primarily from English-speaking countries (UK and USA), limiting cultural generalizability.
     • Scope of scenario design: The selected scenarios do not encompass all possible contexts related to AI behavior.
     • Artificial experimental settings: The experimental population and constraints differ from real-world moral decision-making environments.

  2. Future Directions:

     • Expand research to include a broader range of cultural and social contexts.
     • Explore the potential impact of technological cognition education on reducing AI anthropomorphism tendencies and improving judgment consistency.
     • Further investigate how AI interaction systems can be designed to reduce users' tendencies to anthropomorphize AI.