Principles of Explanation in Human-AI Systems

Algorithm-first approaches to explainability often fail users: The paper finds that many XAI systems are "algorithm-focused; starting and ending with an algorithm that implements a basic untested idea about explainability" without being tested to determine whether they actually help users accomplish their goals, meaning their explainability remains unproven.

Human needs should drive XAI design, not technical capabilities: Rather than building explanations around what algorithms can do, the authors advocate for starting with "human-focused principles for the design, testing, and implementation of XAI systems, and implement algorithms to serve that purpose"—a fundamental shift in how explainable AI should be developed.

Empowering user self-explanation is a key goal: The paper introduces the "Self-Explanation Scorecard" as a tool to help developers understand how to "empower users by enabling self-explanation," suggesting that effective XAI should help people explain AI decisions to themselves rather than simply providing pre-generated explanations.

Four decades of research provides guidance for responsible XAI: The authors draw on "basic concepts that have been used for user-centered XAI systems over the past 40 years of research" to develop "empirically-grounded, user-centered design principles," indicating that responsible AI explainability has established best practices that should be followed.