Designing Responsible AI: Adaptations of UX Practice to Meet Responsible AI Challenges

Title of the Paper

Designing Responsible AI: Adaptations of UX Practice to Meet Responsible AI Challenges

Paper Information

• Subject Areas: Human-Computer Interaction (HCI), User Experience (UX), Responsible Artificial Intelligence (Responsible AI)
• Keywords: Responsible AI, user experience design, industry practices, AI accountability, interview study, responsible AI design methods, prototyping, AI user evaluation

Research Background and Problems

• Problems or Challenges Identified by the Authors:

  1. Current responsible AI (RAI) practices in the technology field have not been adequately studied within the domain of user experience (UX) design.
  2. Most RAI research focuses on data labeling, model evaluation, and developer tools, while overlooking insights into consumer-facing applications of models.
  3. RAI remains a vague area in traditional UX practices, with the roles and responsibilities of UX practitioners yet to be fully clarified and formalized.

• Why This Problem is Important:

  1. The societal risks of AI technology are increasing, necessitating the integration of a responsibility perspective during the design phase.
  2. User experience, as the entry point for user interaction with AI, plays a critical role in promoting the implementation and adoption of RAI technologies.
  3. In the design of emerging large-scale language model applications, there is a significant knowledge and tool gap between the RAI practices of UX practitioners and industry needs.

• Research Motivation and Related Work:

  1. Introduce a research perspective to explore the evolving roles of UX practitioners in RAI and fill the gap in existing RAI research within HCI practices.
  2. Analyze the challenges and opportunities of RAI through the practical experiences of practitioners, further refining industry toolkits and practice guidelines.

Solutions

• Proposed Methods or Solutions:

  1. Conduct multiple semi-structured interviews with UX practitioners and RAI experts to deeply explore practical experiences and design challenges.
  2. Extract three emerging practices to adapt to RAI design needs:
     • Building and strengthening an RAI perspective.
     • Responsible prototyping.
     • Responsible evaluation of AI applications.

• Innovative Aspects of the Solution:

  1. Combines UX design methods with AI ethics design principles, introducing a responsibility-oriented "lens" for participants to consider the potential societal and user impacts of their designs.
  2. Enhances the complexity of user research methods, making them more closely tied to the utility of AI prototypes.
  3. Emphasizes the impact of organizational culture and resource constraints on RAI practices, offering coping strategies and recommendations.

• Implementation Steps and Key Techniques:

  1. Based on research and interviews, comprehensively identify how RAI is embedded in the daily work of UX practitioners.
  2. Introduce the PromptMaker tool to assist practitioners in "test-driving" models, understanding their feasibility and potential risks.
  3. Organize design and evaluation methods to quantify the implicit work and challenges faced by participants in RAI practices.

Research Outcomes

• Specific Achievements:

  1. Summarized three key practices developed by UX practitioners in RAI application design:
     • Building and Strengthening an RAI Perspective: Includes enhancing design teams' sensitivity to RAI issues through educational resources, discussion frameworks, and team culture building.
     • Responsible Prototyping: Identifies and tests potential model risks through methods such as "test-driving" and user input constraints.
     • Responsible Evaluation: Diversifies user groups in user testing while protecting users from potential harmful content.

• Advantages Compared to Existing Solutions:

  1. Addresses the theoretical and abstract nature of RAI by introducing highly actionable UX design methods.
  2. Highlights the evolution of RAI from theory to practice, pointing out potential HCI research opportunities for academia and industry.

• Experimental or Evaluation Results:

  1. Through participatory analysis with UX and RAI practitioners, revealed the implicit responsibility work in AI model design.
  2. Found that the effective practice of RAI frameworks is often constrained by time, manpower, and organizational resources.

• Limitations and Future Directions:

  1. The study is limited to a single company, potentially influenced by the organization's specific culture and processes, requiring broader validation for representativeness.
  2. Interview participants were primarily focused on computer vision and language model domains; future research could expand to other types of AI models.
  3. Explore systematic organizational support for RAI work and optimize UX practitioner education and RAI tool design.