Expanding Explainability: Towards Social Transparency in AI systems

Document Title

Expanding Explainability: Towards Social Transparency in AI Systems

Document Information

• Subject Area: Explainable Artificial Intelligence (XAI)
• Keywords: Explainable AI, Social Transparency, Human-Computer Interaction, Explainability, Artificial Intelligence, Sociotechnical, Organizational Context.

Research Background and Issues

• Identified Problems and Challenges:

  • Current XAI methods overly emphasize algorithmic transparency and fail to effectively incorporate the social and organizational needs of human users regarding explainability.
  • Human users' actual needs for AI explanations in high-stakes decisions (e.g., healthcare, employment, criminal justice) remain unmet.
  • Existing XAI technologies have limited effectiveness in enhancing user trust and task performance and may even pose potential risks.

• Importance of the Problem: With the widespread application of AI technologies in critical decision-making domains, ineffective or inadequate explanation methods may lead to misunderstandings, trust imbalances, or erroneous decisions.

• Research Motivation: The authors argue that these shortcomings stem from the technocentric tendencies of XAI methods, which neglect the social context of explanations. The research aims to address this gap by introducing the concept of "Social Transparency" (ST).

Solution

• Proposed Method or Solution:

  • Introduce the concept of "Social Transparency" (ST) as an explainability approach that incorporates social and organizational contexts.
  • Explore the impact of visualizing "Who, What, When, Why" (4W) information on user decision-making and explainability through the design of AI systems with ST features.

• Innovative Aspects of the Solution:

  • Incorporate social and organizational contexts into AI explanations, going beyond mere algorithmic transparency.
  • Design specific interface elements based on the "4W" framework, providing decision contexts through user interaction histories.
  • Employ a projection-based design approach to conceptually explore how social transparency influences human-AI decision-making.

• Implementation Steps and Key Techniques:

  • Utilize Scenario-Based Design (SBD).
  • Develop prototype designs based on sales scenarios, integrating multiple ST feature modules (e.g., visualization of 4W information).
  • Collect feedback through interviews and user experiments to investigate the potential effects of ST.
  • Employ mixed data analysis methods, including qualitative analysis and statistical data, to comprehensively understand changes in user behavior and perceptions.

Research Findings

• Specific Findings Achieved:

  • ST significantly impacts users' trust in AI, decision confidence, and optimized actions, addressing the shortcomings of purely technical transparency in explainability.
  • Clarified the visualization content and impact of ST at three levels:
    1. Technical Level: Provide AI's past behavior trajectories and user interaction records to calibrate user trust in AI.
    2. Decision Level: Offer social validation of the context surrounding AI-generated decisions, supporting users in reasonably rejecting AI recommendations.
    3. Organizational Level: Provide meta-knowledge (e.g., identifying experts) and group norms to enhance the efficiency of collective actions.
  • Proposed specific interface functionality suggestions based on the "4W" design elements (Who, What, When, Why).
  • Demonstrated the cross-domain generalizability of ST, with particularly strong demand in sales, security, and healthcare domains.

• Advantages Over Existing Solutions:

  • Offers a context-rich decision explanation framework, effectively overcoming the limitations of relying solely on technical transparency.
  • Combines users' social validation with algorithmic explanations, empowering users to construct a holistic understanding of AI.

• Experimental and Evaluation Results:

  • In user experiments, 26 out of 29 participants adjusted their decisions based on AI recommendations and reported increased decision confidence.
  • Functionality rankings indicated that "What" and "Why" information most effectively helped users perceive the technical and decision-making context.

• Limitations and Future Directions:

  • Limitations: The current study is based on scenario design and remains exploratory, lacking validation of long-term effectiveness in real-world contexts. Implementing ST may face challenges related to privacy and bias.
  • Future Directions:
    • Broader investigation of ST's applicability in other domains.
    • Explore methods for feeding social data back into AI models to improve performance and explainability.
    • Address challenges of information overload from ST features, such as utilizing NLP tools to process review data.
    • Extend research on the long-term impacts of ST (e.g., user trust) and develop methods to mitigate potential negative effects.

Summary

This study introduces and explores the novel concept of "Social Transparency" (ST), expanding the design space of XAI by integrating explainability into social and organizational contexts. Through design-driven experiments and interdisciplinary sociotechnical perspectives, this research takes a first step towards achieving socially contextualized XAI.