Blaming Humans and Machines: What Shapes People's Reactions to Algorithmic Harm

Title of the Paper

Blaming Humans and Machines: What Shapes People’s Reactions to Algorithmic Harm

Paper Information

• Research Domain: Human-Computer Interaction, Algorithm Ethics, Artificial Intelligence
• Keywords: Artificial Intelligence, Responsibility Attribution, Ethics, Explainability, Decision-making, Algorithmic Bias, Fairness, Human-Computer Interaction

Research Background and Problem

• Issues and Challenges:
  • An increasing number of AI systems are being deployed in high-risk scenarios (e.g., medical diagnosis, recruitment), raising concerns about their potential ethical implications.
  • Existing research has revealed that AI systems can exhibit biases across gender, race, and other dimensions, sometimes leading to direct harm (e.g., fatal accidents involving autonomous vehicles).
  • People's reactions to algorithmic harm are inconsistent—some show tolerance, while others respond with anger or even destructive behavior.
• Significance:
  • Understanding how people react to algorithmic harm and attribute responsibility is crucial for managing, developing, and promoting AI systems.
  • The issue touches upon how AI integrates into societal moral frameworks and the design of corresponding public policies.
• Research Motivation and Related Work:
  • Philosophical perspective: Debates exist regarding whether machines possess attributes that warrant blame (e.g., self-awareness, sense of responsibility).
  • Empirical perspective: Previous studies suggest that people attribute responsibility to AI and robots differently than to humans, but systematic exploration of how responsibility is attributed is lacking.

Solution

• Methodology and Innovation:
  • Propose a multi-experiment design to explore how various factors (e.g., AI explainability, harm perception, fairness) influence people's attribution of responsibility to AI systems, their developers, and users.
  • Innovations include employing three experimental groups (total participants: 1153) to deeply compare the driving mechanisms behind different attribution behaviors.
• Implementation Steps:
  1. Experimental Design:
     • Experiment 1: Test the impact of AI explainability on responsibility attribution using a recruitment scenario, with conditions of no explanation, experience-based explanation, and discriminatory explanation.
     • Experiment 2: Introduce the dimension of "perceived harm severity" using a medical scenario, observing differences in attribution behavior under non-lethal and lethal consequences.
     • Experiment 3: Explore whether AI autonomy affects attribution behavior, further distinguishing between lawful and unlawful discriminatory explanations.
  2. Data Collection and Evaluation:
     • Use questionnaires to record participants' attitudes toward responsibility attribution for AI systems and related human actors, combined with perceptions of fairness and explainability.
  3. Meta-analysis:
     • Conduct an internal meta-analysis of the results from the three experiments to validate consistency.

Research Findings

• Key Discoveries:

  • AI System Explainability:
    • Providing explanations alone does not significantly influence people's attribution of responsibility to AI systems or their developers and users.
    • Explanations involving discriminatory factors significantly increase responsibility attributed to developers and users, but have limited impact on responsibility attributed to the AI system itself.
  • Perceived Harm and Fairness:
    • Perceived harm drives responsibility attribution more toward AI users (e.g., hospitals).
    • Fairness primarily affects the degree of responsibility attributed to developers.
    • Responsibility attributed to AI systems is independent of fairness, which differs significantly from human attribution patterns.
  • Attitudes Toward AI:
    • Those who believe AI can be "blamed" attribute higher levels of responsibility to AI systems.
    • Conversely, those who deny AI's capacity for blame assign limited responsibility to AI, while attributing more blame to human developers and users.
  • Relationship Between AI Responsibility and Developer/User Responsibility:
    • Attributing responsibility to AI systems does not significantly reduce responsibility attributed to developers or users.

• Policy and Design Implications:

  • The design and deployment of new technologies should consider public expectations regarding responsibility distribution.
  • In high-risk scenarios, developers should bear primary responsibility, while users need to adopt systems cautiously and mitigate potential risks.
  • Public policy could consider introducing joint and proportional responsibility-sharing models.

• Limitations and Future Directions:

  • The experiments were based on virtual scenarios, which may not fully capture the complexity of real-world contexts.
  • Beyond "fairness" and "perceived harm," future research could explore other ethical dimensions (e.g., intentionality) that influence attribution behaviors.
  • Further investigation is recommended into how different cultural backgrounds, experiences, and demographics shape reactions to algorithmic harm.