Mental Models of Autonomy and Sentience Shape Reactions to AI

Paper Title

Mental Models of Autonomy and Sentience Shape Reactions to AI

Publication Info

• Topic area: Human-computer interaction (HCI) and mental models in AI perception.
• Keywords: AI autonomy, AI sentience, mental models, human-AI interaction, mind perception, moral consideration, perceived threat, AI governance, anthropomorphism, digital minds.

Background and Problem

• Problem / challenge: Mental models of AI autonomy and sentience are often conflated in public narratives and research, leading to a lack of systematic understanding of their distinct and combined effects on human reactions to AI systems.
• Significance: Understanding these mental models is critical for improving AI design, fostering prosocial human-AI interactions, and informing governance and policy decisions in the context of increasingly anthropomorphized AI systems.
• Motivation and related work: Prior research has explored mental models of AI as opaque, emotionless, or autonomous, but has not disentangled autonomy and sentience. These models influence trust, moral consideration, and threat perception, yet their independent and interactive effects remain underexplored.

Solution

• Proposed approach: The study disentangles mental models of autonomy and sentience through a series of pilot studies and experiments, using vignettes to activate these models and measure their effects on mind perception, moral consideration, perceived threat, and policy preferences.
• Novelty:
  1. Systematic disentanglement of autonomy and sentience in mental models of AI.
  2. Empirical evidence from four experiments and a meta-analysis on their distinct and combined effects.
  3. Insights into the downstream impacts of these mental models on human-AI interaction, design, and policy.
  4. Practical recommendations for designing AI systems and fostering AI literacy.
• Procedure and key techniques:
  • Conducted three pilot studies (N = 374) to map mental model content.
  • Performed four vignette-based experiments (N = 2,702) manipulating autonomy and sentience in a hypothetical AI assistant.
  • Measured outcomes such as mind perception, moral consideration, perceived threat, and support for AI regulation.
  • Conducted a meta-analysis to compare the magnitude of autonomy and sentience effects.

Results

• Concrete findings:
  • Sentience had a larger effect than autonomy on mind perception (Cohen’s d = 0.92 vs. 0.72).
  • Autonomy increased perceived harm but not policy support for AI regulation.
  • Combined autonomy and sentience produced the highest mind perception, while their absence produced the lowest.
  • Sentience increased moral consideration more than autonomy.
• Advantage over baselines:
  • Disentangling autonomy and sentience revealed nuanced effects not captured by treating AI as a monolithic category.
  • Sentience was shown to anchor perceptions of autonomy, highlighting its foundational role in mind perception.
• Experiments / evaluation:
  • Experiment 1: Autonomy increased mind perception, moral consideration, and perceived harm (N = 254).
  • Experiment 2: Sentience increased mind perception, moral consideration, and policy support (N = 256).
  • Experiment 3: Combined autonomy and sentience produced the highest mind perception and moral consideration (N = 741).
  • Experiment 4: Removing explicit labels confirmed robust autonomy and sentience effects (N = 1,451).
  • Meta-analysis confirmed sentience effects were larger than autonomy effects across all experiments.
• Limitations and future work:
  • Hypothetical nature of sentience in AI limits real-world applicability.
  • Need for further exploration of interactions between autonomy, sentience, and other mental models (e.g., anthropomorphism).
  • Future studies should examine long-term mental model updates and their effects on human-AI interaction.

Summary

This study disentangles mental models of autonomy and sentience in AI, demonstrating their distinct and combined effects on mind perception, moral consideration, perceived threat, and policy preferences. Sentience had a stronger influence than autonomy, anchoring perceptions of autonomy and driving moral consideration. Autonomy increased perceived harm but not policy support. These findings inform AI design, emphasizing the need for targeted tuning of autonomy and sentience cues to align with user expectations while avoiding misleading representations. The results also highlight the importance of AI literacy and nuanced governance to address public perceptions and foster prosocial human-AI interactions.