Designerly Understanding: Information Needs for Model Transparency to Support Design Ideation for AI-Powered User Experience

Title of the Paper

Designerly Understanding: Information Needs for Model Transparency to Support Design Ideation for AI-Powered User Experience

Paper Information

• Subject Area: Human-Computer Interaction (HCI), AI Design
• Keywords: AI Design, Model Transparency, AI Documentation, Explainability, Pre-trained Models

Research Background and Problem

• Challenges Identified by the Authors:

  1. UX designers face complexity and professional barriers when designing AI-driven systems, particularly in understanding AI technologies like pre-trained language models as design materials.
  2. The lack of methods supporting UX-led AI product innovation results in product failures and unintended societal consequences.
  3. While the widespread use of pre-trained models lowers technical barriers, it increases accountability, necessitating responsible AI design practices.

• Importance of the Problem:

  • AI technologies are often disconnected from user needs. To fully leverage AI's potential, it is essential to identify scenarios where it can solve user problems and optimize model behavior based on user preferences. These tasks are central to UX design but are currently under-supported.

• Research Motivation and Related Work:

  • Applying AI transparency frameworks to the design process to support designers in developing a "designerly understanding" of AI models and exploring the practicality and limitations of various model documentation frameworks.

Solution

• Proposed Solution:

  • Investigate how UX designers use transparent model reporting frameworks to develop a designerly understanding of pre-trained models through interviews and scenario-based design tasks.

• Innovations:

  1. Bridging AI design and AI transparency research to help UX designers understand model limitations and their potential impact on users.
  2. Proposing four design goals and detailing the information designers need for each goal.
  3. Highlighting the critical role of UX designers in responsible AI, where designerly understanding aids in evaluating model limitations and suggesting improvements.

• Implementation Steps and Key Techniques:

  1. Using example documentation and input-output comparison data from AI services as design probes.
  2. Introducing interactive design tasks using real-world scenarios based on pre-trained text summarization models, enabling participants to design new product features to address information misunderstanding issues.
  3. Exploring designers' information needs, usage strategies, and design goals through interviews.

Research Findings

• Specific Findings:

  • Based on interviews with 23 UX design practitioners:

    1. Designers face diverse challenges, such as the lack of support for evaluating model applicability and limited opportunities for collaboration with data scientists.
    2. Current transparency documentation is useful but has gaps, requiring additional information to support design tasks.

  • Four common design goals for designers were identified:

    1. Evaluating and eliminating risky design ideas during divergent-to-convergent design thinking.
    2. Creating conditional designs to meet the needs of different user scenarios and mitigate AI system uncertainty.
    3. Providing end-users with transparent information about AI models.
    4. Collaborating and negotiating with teams to advocate for better designs on behalf of users.

• Advantages:

  • Helps UX designers better understand AI models, driving user-centered design.
  • Provides a framework enabling designers to predict model limitations and risks, strengthening responsible AI practices.

• Experimental or Evaluation Results:

  • The information provided is insufficient to meet all designers' needs, but transparency documentation allows designers to better extract information and conceptualize design solutions.
  • Some designers gained deeper insights into model characteristics through input-output examples, document reviews, and speculations on model limitations, but lacked more detailed explanatory information.

• Limitations and Future Directions:

  • Limitations:
    1. The specific pre-trained models and design scenarios used may affect the generalizability of the experimental results.
    2. Participants were UX designers from multiple tech companies, but most were from a single company, potentially limiting sample representativeness.
  • Future Directions:
    1. Explore more suitable information interaction tools for AI designers (e.g., interactive model explanation documentation).
    2. Provide support for designers to explore dependent variables, helping them test how module characteristics impact UX.
    3. Develop guiding frameworks to support UX designers in actively participating in the lifecycle of responsible AI development.

Summary

This study demonstrates the potential of leveraging model transparency to support UX designers in understanding pre-trained models and generating design ideas. It further emphasizes the role of designers in responsible AI, highlighting their information needs that extend beyond current transparency frameworks to include deeper model and user-related contextual knowledge. This provides a solid theoretical foundation for the development of future tools and frameworks aimed at enhancing the utility of transparency frameworks and fostering interdisciplinary collaboration.