Which Artificial Intelligences Do People Care About Most? A Conjoint Experiment on Moral Consideration

Title of the Paper

Which Artificial Intelligences Do People Care About Most? A Conjoint Experiment on Moral Consideration

Paper Information

• Subject Area: Human-Computer Interaction and Moral Consideration
• Keywords: Morality, Prosociality, Anthropomorphism, Human Traits, Human-Computer Interaction, Collaboration and Social Computing, Conjoint Experiment

Research Background and Questions

• Research Questions and Challenges:
  • In increasingly complex AI systems, it remains unclear how different characteristics influence people's moral consideration of them.
  • Do people think it is morally wrong to harm AI? What characteristics lead AI to be included in a higher range of moral consideration?
• Research Importance:
  • Highly realistic AIs (e.g., social robots, chatbots) have triggered complex moral reactions, such as sympathy, protective behavior, and even ethical guilt when these robots are harmed.
  • For future AI design, understanding which characteristics drive moral reactions is crucial for optimizing human-computer interaction design.
• Research Motivation and Related Work:
  • Existing studies often focus on single characteristics of AI (e.g., autonomy, emotional expression) but lack systematic analysis of their relative importance.
  • The authors conducted an online conjoint experiment to systematically study the combined effects of 11 different characteristics on moral consideration of AI.

Solution

• Research Method Overview:
  • The authors designed a conjoint experiment in which 1,163 participants compared and evaluated descriptions of fictional AIs to assess which characteristics led to stronger moral consideration.
  • The study examined 11 AI characteristics: autonomy, physical form, complexity, cooperativeness, harm avoidance, emotional expression, emotion recognition, intelligence, language ability, moral judgment, and purpose.
• Key Techniques and Design:
  • Experimental Design:
    • Randomized design of partial feature configurations to minimize participants' cognitive load.
    • Linear regression models were used to estimate the Average Marginal Component Effect (AMCE) of each characteristic.
  • Basis for Characteristic Selection: A comprehensive literature review and pilot testing led to the final selection of 11 highly relevant characteristics.
  • Evaluation Dimensions: Measured the impact of described characteristics (e.g., "not at all," "to some extent," "to a great extent") on moral wrongness ratings of harming AI.
• Innovations:
  • The first experimental study to systematically evaluate the relative impact of multiple characteristics on moral consideration of AI.
  • Integration of data across different social psychology domains provides strong practical guidance for HCI research.

Research Findings

• Key Discoveries:
  • All 11 characteristics significantly increased moral consideration ratings.
  • The characteristics with the greatest impact on moral consideration were:
    1. Moral judgment capability.
    2. Emotional expression (e.g., displaying emotions).
    3. Emotion recognition (perceiving others' emotions).
    4. Cooperativeness (friendly collaboration with humans).
    5. Human-like physical form.
  • Secondary influential characteristics included autonomy, complexity, harm avoidance, language ability, and social-purpose goals.
• Experimental Results Data:
  • AMCE analysis showed that moral judgment capability and emotion-related characteristics increased moral consideration ratings by over 20%.
  • Functional characteristics (e.g., task complexity or independence) had weaker effects, with impact ranges between 6%-9%.
• Advantages Compared to Existing Approaches:
  • Unlike prior studies focusing on single-dimensional variables, this experiment quantified the relative influence of multiple characteristic combinations, offering a holistic perspective for academia and design practice.
  • Directly supports the importance of prosocial behavior in AI design.
• Limitations and Future Directions:
  1. Participants were from the U.S., limiting the global generalizability of the sample.
  2. The study only explored hypothetical preferences in specific scenarios, requiring validation in real-time interactive environments for external validity.
  3. The current research measured only a single dimension of moral consideration; future studies should incorporate multi-dimensional measurements (e.g., legal attributes, ethical constraints) in more complex behavioral contexts.
  4. Differentiating human-AI cognitive and psychological responses across cultural backgrounds.

Conclusion

This study comprehensively demonstrates the multidimensional performance of artificial intelligence in the domain of moral consideration, revealing the central role of prosociality and anthropomorphic characteristics in shaping users' moral reactions to AI. For AI interaction designers, this implies that to build more trustworthy and efficient AI systems in the future, greater emphasis should be placed on social and emotional interaction functions, while also balancing the ethical and resource consumption challenges such designs may pose.