Rethinking User Empowerment in AI Recommender System: Innovating Transparent and Controllable Interfaces

Paper Title

Rethinking User Empowerment in AI Recommender System: Innovating Transparent and Controllable Interfaces

Publication Info

• Topic area: Enhancing user agency in AI-driven recommender systems through transparency and control.
• Keywords: AI recommender systems, user agency, transparency, control, filter bubbles, personalization, human-AI interaction, user interface design, algorithmic fairness, explainable AI.

Background and Problem

• Problem / challenge: Current AI-driven recommender systems (RS) often function as opaque "black boxes," leading to two key asymmetries:
  1. Information asymmetry: Users cannot see how recommendations are generated.
  2. Power asymmetry: Users have limited ability to influence or control recommendations. These asymmetries result in issues like filter bubbles, misaligned personalization, and reduced trust.
• Significance: Addressing these asymmetries is critical for fostering user autonomy, improving trust, and mitigating negative consequences such as manipulation, information isolation, and reduced system adoption.
• Motivation and related work: Prior work has explored transparency (e.g., explainable AI) and user control mechanisms (e.g., feedback systems), but these solutions are often partial, backend-centric, or overly rigid. This paper builds on these gaps by exploring interface-level interventions that enhance user agency and align personalization with user preferences.

Solution

• Proposed approach: A provotype (provocative prototype) was designed to introduce new interface features for managing data use, exploring diverse content, and configuring context-based recommendation modes.
• Novelty:
  1. Development of actionable transparency features to address information asymmetry.
  2. Introduction of user-driven content discovery mechanisms to counter filter bubbles.
  3. Creation of scenario-based "Personal Modes" for dynamic, context-aware personalization.
  4. Empirical insights into user perceptions of transparency, control, and trust in RS.
• Procedure and key techniques:
  1. Design of a provotype with interactive features such as Data Management, Content Discovery, and Personal Modes.
  2. User study with 19 participants, including walkthroughs, semi-structured interviews, and surveys to evaluate the provotype.
  3. Mixed-methods analysis combining quantitative comparisons and qualitative thematic analysis.

Results

• Concrete findings:
  • Significant increases in perceived transparency (2.95 to 4.63, p < .001) and control (2.77 to 4.51, p < .001) compared to existing RS.
  • Features like Adventure Content and Personal Modes were highly favored for breaking monotony and adapting to user contexts.
  • Data management features (e.g., Data Type Weighting, Data Aging) were seen as empowering but raised concerns about complexity and emotional discomfort.
• Advantage over baselines:
  • The provotype shifted RS from opaque, prescriptive systems to user-driven tools, fostering greater trust and engagement.
  • Participants reported stronger feelings of safety, empowerment, and willingness to engage compared to existing RS.
• Experiments / evaluation:
  • 19 participants (mean age = 28.11) interacted with the provotype in a 75-minute session, followed by surveys and interviews.
  • Metrics included perceived transparency, control, and trust, analyzed through Wilcoxon signed-rank tests and thematic coding.
• Limitations and future work:
  • The provotype was non-executable and tested only in a desktop format, limiting real-world applicability.
  • The sample was skewed toward young, tech-savvy participants, raising questions about generalizability.
  • Future work should test real-world implementations, explore broader user demographics, and investigate unresolved socio-technical tensions.

Summary

This paper addresses the asymmetries in AI-driven recommender systems by designing a provotype with features that enhance transparency, control, and user agency. The study demonstrated significant improvements in perceived transparency, control, and trust, with features like Adventure Content and Personal Modes being particularly well-received. However, challenges such as balancing complexity, addressing privacy concerns, and ensuring usability remain. The findings provide actionable insights for designing RS that empower users while maintaining algorithmic efficiency, offering a roadmap for more agency-oriented AI systems.