Demystifying XAI: Requirements for Understandable XAI Explanations

Area of Science:

• Human-Computer Interaction
• Artificial Intelligence
• Healthcare Informatics

Background:

• Explainable Artificial Intelligence (XAI) methods require usability assessments tailored for non-AI experts.
• Healthcare professionals need transparent and understandable AI explanations for effective clinical decision-making.
• Current XAI approaches may not adequately address the cognitive and trust needs of end-users in high-stakes domains.

Purpose of the Study:

• To establish clear requirements for assessing the usability of XAI methods.
• To guide the design of XAI explanations that are comprehensible and trustworthy for non-expert users.
• To facilitate seamless human-AI collaboration in fields like healthcare.

Main Methods:

• Synthesis of existing literature on XAI usability.
• Analysis of empirical findings related to user interaction with AI explanations.
• Identification of key factors influencing user comprehension, trust, and performance.

Main Results:

• Optimal cognitive load, task performance, and task time are critical for XAI usability.
• Tailoring explanations to user expertise, integrating domain knowledge, and using non-propositional representations enhance comprehension.
• Relevance, accuracy, and truthfulness are essential for building user trust in XAI systems.

Conclusions:

• Effective XAI explanations must be transparent, user-friendly, and context-aware.
• Practical guidelines are provided for designing usable XAI, particularly for healthcare applications.
• This work contributes essential requirements for advancing human-AI collaboration through improved XAI design.