Solving Separation-of-Concerns Problems in Collaborative Design of Human-AI Systems through Leaky Abstractions

Title of the Paper

Solving Separation-of-Concerns Problems in Collaborative Design of Human-AI Systems through Leaky Abstractions

Paper Information

• Subject Area: Artificial Intelligence (AI) and User Experience Design (UX)
• Keywords: AI systems, user experience design, AI applications, industry practices, design processes, collaboration, knowledge boundaries, information leakage, software development, interdisciplinary work

Research Background and Problem Statement

• Identified Issues or Challenges:

  • In traditional software development, "Separation of Concerns" (SoC) is an effective design principle aimed at improving efficiency by clearly defining responsibilities in design and implementation.
  • However, in the design of human-centered artificial intelligence (Human-AI, HAI) systems, the isolation caused by SoC hinders effective collaboration between designers and engineers.
  • HAI systems involve the dynamic characteristics of AI models and high levels of uncertainty, complicating the collaboration between design and engineering teams.
  • Challenges such as designers' limited understanding of AI technologies and engineers' neglect of user experience design requirements result in teams failing to deliver AI solutions that meet human needs.

• Significance:

  • Key areas of current research in AI system design include human-computer interaction, task automation, design explainability, and data privacy.
  • More effective interdisciplinary collaboration can help foster high-quality AI applications while addressing issues such as model bias, user experience challenges, and ensuring the ethical and societal value of AI designs.

• Research Motivation and Related Work:

  • Turing Award winners and researchers in related fields have emphasized the importance of integrating AI model design with human data needs.
  • While guiding principles for HAI design have been proposed in the industry, specific collaborative methods for interdisciplinary team design are lacking.
  • Current research primarily focuses on the challenges faced by designers and data scientists individually, neglecting the actual processes of collaboration between designers and engineers.

Proposed Solution

• Proposed Solution:

  • This study proposes the use of the "Leaky Abstractions" design approach to overcome the barriers of Separation of Concerns in traditional software development.
  • By sharing low-level design details and implementation information, the collaboration between designers and engineers can be improved.
  • The study develops an interdisciplinary iterative design process combining Delayed Specification and Constant Evaluation.

• Innovations:

  • Introduced the novel design method of "Leaky Abstractions," emphasizing the breaking of knowledge boundaries through interdisciplinary sharing of low-level design and engineering details.
  • Defined a component model for HAI design guidelines, breaking down the design process into specific submodules: user mental models, user interface design, AI model design, and data training requirements.
  • Constructed a theoretical framework for interdisciplinary collaboration, extending traditional design communication tools and methods.

• Implementation Steps and Techniques:

  1. Analysis of HAI Design Guidelines:
     • Collected 280 HAI design guidelines from major industry sources and summarized key components of HAI design through association graph analysis.
  2. Industry Practice Research:
     • Conducted interviews with 21 industry practitioners from various companies to explore their collaborative methods and challenges in HAI design.
     • Collected real-world cases to validate the effectiveness and practical application scenarios of "Leaky Abstractions."
  3. Component Model and Design Prototypes:
     • Developed interaction models and components for user-AI interaction, designing "gray-box" prototypes for early testing.
     • Facilitated cross-role collaboration through engineering documentation, dataset specifications, and UI annotations.

Research Outcomes

• Specific Outcomes:

  • Developed a component model for HAI design, including four major modules: user mental models (task models, expectation models, interaction models), user interface design (inputs and outputs, explainability, failure and handoff), AI models (performance evaluation, user learning), and data training requirements.
  • Identified key barriers to collaboration between designers and engineers, such as knowledge blind spots and misaligned timelines.
  • Proposed practical methods to enhance team collaboration by sharing low-level details (e.g., data labeling, model assumptions, prototype outputs) through "Leaky Abstractions."

• Advantages and Comparisons:

  • Compared to traditional Separation of Concerns practices, "Leaky Abstractions" addresses information asymmetry in team collaboration, encouraging innovation and flexible design.
  • Delayed specification reduces costly late-stage changes caused by premature decisions.

• Experimental or Evaluation Results:

  • Interview data revealed that engineers and designers significantly reduced information gaps in collaboration by using "Leaky Abstractions" to share design prototypes, data analysis tools, and task scenarios.
  • Provided examples, such as designers participating in the creation of data labeling rules and engineers sharing ML system performance visualizations to optimize user interface design.

• Limitations and Future Directions:

  • Due to the limited scope of interviews and demonstrations, this study may not fully cover the diversity of fields, tasks, and user groups.
  • Future research should focus on integrating "Delayed Specification" with agile development practices and systematizing standardized tools.
  • Further exploration is needed on how socio-technical practices can support responsible AI design, particularly regarding fairness metrics and user data privacy.

This study provides new perspectives and practical directions for human-AI collaborative design while uncovering potential challenges and solutions in current industry practices.