One AI Does Not Fit All: A Cluster Analysis of the Laypeople’s Perception of AI Roles

Document Title

One AI Does Not Fit All: A Cluster Analysis of Laypeople’s Perception of AI Roles

Document Information

• Subject Area: AI Cognition and Human-Computer Interaction
• Keywords: Artificial Intelligence (AI), Human-AI Interaction (HAII), Human Involvement, AI Autonomy, Trustworthiness, Attitudes, Social Acceptance, Cluster Analysis

Research Background and Problem

• Issues and Challenges:

  • Current research on Artificial Intelligence (AI) either treats AI as a singular entity or focuses on specific application scenarios, both of which have limitations.
  • The diverse roles of AI lead to variations in how the public evaluates different types of AI. Ignoring this diversity may result in insufficient ecological validity.
  • How can research balance generalizability with specific scenarios? How can a unified design framework be created across domains?

• Significance:

  • With the proliferation of AI in daily life, such as virtual assistants and autonomous driving systems, understanding and optimizing human-AI interaction is increasingly critical.
  • Understanding public perceptions of AI's diverse roles can better guide design and enhance AI's social acceptance.

• Research Motivation and Related Work:

  • Summarized and classified AI's various functional roles, such as assistant, tool, servant, etc., addressing the gap in data-driven classification through theoretical innovation.
  • Leveraged existing theories on AI's perceived mind, perceived control, and moral agency to distill the core dimensions of human perception of AI.

Solution

• Methods or Solutions:

  • Employed computational methods to classify AI roles, using Principal Component Analysis (PCA) to extract the core dimensions of human perception: "AI Autonomy" and "Human Involvement."
  • Based on these dimensions, conducted a cluster analysis of 10 common AI roles in daily life, resulting in four main categories: AI Tools, AI Servants, AI Assistants, and AI Mediators.

• Innovations:

  • Proposed a classification framework that combines theoretical and data-driven approaches, balancing generalizability and ecological validity in specific scenarios.
  • Emphasized the balance between AI autonomy and human involvement, aiding in the design of AI systems that are both trustworthy and appropriately reliable.

• Key Techniques and Steps:

  • Statistical Analysis: PCA was used to reduce theoretical dimensions, summarizing six variables (sub-dimensions of perceived mind, perceived control, and moral agency) into two.
  • Cluster Analysis: Standardized data and k-means clustering were applied to identify four AI role categories.
  • Survey Method: Conducted an online survey with 727 participants to measure their perceptions of different AI roles.

Research Findings

• Specific Findings:

  • Identified four AI role categories: Tools (low autonomy, low human involvement), Servants (high human involvement, low autonomy), Assistants (low human involvement, high autonomy), and Mediators (high autonomy, high human involvement).
  • AI Mediators received the most positive evaluations, with the highest trustworthiness, attitudes, and social acceptance; AI Tools received the most negative evaluations.
  • Demographic variables such as gender, race, age, and education level significantly influenced perceptions of AI.

• Comparative Advantages Over Existing Solutions:

  • This study integrates theoretical and data-driven analyses, providing valuable references for classifying AI roles and optimizing their design.
  • The cluster analysis results have practical applications, helping to enhance user acceptance and trust in AI.

• Experimental or Evaluation Results:

  • The two dimensions extracted by PCA explained 50% of the total variance in user perceptions.
  • People preferred AI roles with high autonomy but also allowing human involvement, such as "Mediators," supporting the importance of human-AI collaboration.

• Limitations and Future Directions:

  • Limitations: The 10 AI roles studied do not cover all possible AI application scenarios, such as judges or artists.
  • Future Directions: Expand to more AI roles and task contexts; explore the impact of other dimensions (e.g., tactility or risk levels) on AI role classification.

Summary and Application Recommendations

• Core Insights:

  • Balancing AI autonomy and human involvement is crucial for enhancing user acceptance of AI.
  • AI design should avoid simply increasing autonomy or control but instead promote appropriate trust through collaborative design.

• Practical Applications:

  • Optimize AI tools, such as customer service bots, by enhancing autonomy and user control.
  • Improve the explainability and transparency of AI assistants to increase user understanding of their decision-making processes.
  • Design integrated mediator systems that combine user and AI control, avoiding mismatched trust.