XAIR: A Framework of Explainable AI in Augmented Reality

Title of the Paper

XAIR: A Framework of Explainable AI in Augmented Reality

Bibliographic Information

• Research Domain: Explainable Artificial Intelligence (XAI) and Augmented Reality (AR)
• Keywords: Explainable Artificial Intelligence (XAI), Augmented Reality (AR), design framework, user experience, intelligent systems, recommendation systems, user trust, model transparency

Research Background and Problem Statement

• Identified Issues or Challenges:

  • As Augmented Reality (AR) technology increasingly integrates with Artificial Intelligence (AI), providing clear and transparent AI model explanations for non-expert users becomes more critical. However, designing effective XAI experiences remains challenging.
  • The unique characteristics of AR (e.g., real-time perception of user state and environment) pose distinct challenges in designing an AR-specific XAI framework.
  • Existing XAI frameworks primarily target developers or experts, lacking design guidance tailored for general users and AR-specific scenarios.

• Significance:

  • Enhancing non-expert users' trust in AR systems, reducing confusion, and improving transparency of experiences are essential.
  • The widespread adoption of AR requires more interpretable intelligent services to build user trust and alleviate concerns about errors and data privacy.

• Research Motivation and Related Work:

  • In recent years, industrial and academic research on XAI has grown rapidly, with most efforts focused on serving AI model developers and industry experts, while paying little attention to XAI design for general users or AR scenarios.
  • The authors aim to address this gap by designing a comprehensive XAI framework tailored for AR.

Solution

• Method and Framework:

  • The authors propose a design framework named XAIR to guide the delivery of effective AI output explanations in AR scenarios.
  • The framework identifies three key design questions:
    1. When to explain: Includes the availability and triggering mechanisms for explanations.
    2. What to explain: Covers the types and levels of detail in the content.
    3. How to explain: Involves visualization or audio representation methods and interaction modes.
  • The framework design process involved the following steps:
    1. Multidisciplinary literature review to define the design problem space and dimensions.
    2. User survey with over 500 participants to explore preferences for AR explanation design.
    3. Multiple rounds of workshops with 12 experts to integrate domain-specific insights.
    4. Dual validation of the framework with designers and general users.

• Key Technologies:

  • User needs analysis methods (user surveys and semi-structured interviews).
  • Integration of AR’s user context inference capabilities with multimodal presentation features.
  • Documented explanation types (Why/Why-Not, How, Certainty, etc.).

Research Outcomes

• Specific Results:

  • Designed and validated the XAIR framework, proposing 8 specific design principles for the design process.
  • Summarized 5 key factors influencing AR XAI design: user state, contextual information, system goals, user goals, and user individual characteristics.
  • Developed two example designs for everyday scenarios under the framework’s guidance (e.g., running navigation recommendations and plant care suggestions).

• Advantages Compared to Existing Solutions:

  • Provides specific explainability design solutions tailored to AR environments (e.g., scene adaptability and spatial awareness).
  • Addresses the needs and cognitive load of general users, offering default options and detailed experience choices.

• Evaluation and Experimental Results:

  • In design experiments, 10 designers using XAIR produced more consistent and systematic designs (rated 7.9/10 for creative support provided by the framework).
  • Real-time AR systems developed based on designer recommendations were tested. 12 general users rated the systems highly usable (SUS scores of 86 and 80, respectively), with AI explanations significantly improving transparency and trust.

• Limitations and Future Directions:

  • Limitations:
    • The framework does not address certain potential dimensions (e.g., "who is being explained to" or "where the explanation occurs").
    • Experimental scenarios focused on recommendation systems, without covering broader application contexts.
    • The design relies entirely on human intuition, leaving the automation of judgment unexplored.
  • Future Directions:
    • Develop a fully automated design recommendation tool to translate the framework into algorithmic decision-making processes.
    • Expand the framework to support new user groups such as developers and domain experts.
    • Investigate long-term user-system co-learning processes to further enhance research on user-customizable XAI features.

Through the XAIR framework, the authors effectively integrate multidisciplinary knowledge to meet the design needs of XAI in augmented reality scenarios, providing designers with a systematic tool.