Information technology — Artificial intelligence — Objectives and approaches for explainability and interpretability of machine learning (ML) models and artificial intelligence (AI) systems
Introduction: When AI systems are used to help make decisions that affect people’s lives, it is important that people understand how those decisions are made. Achieving useful explanations of the behaviour of AI systems and their components is a complex task. Industry and academia are actively exploring emerging methods for enabling explainability, as well as scenarios and reasons why explainability can be required. Due to the multitude of stakeholders and communities contributing to this effort, the field is suffering from a certain terminological inconsistency. Most notably, the methods to provide such explanations of the behaviour of an AI system are discussed under the banner of “explainability”, “interpretability”, (sometimes even other terms like “transparency”), raising the question of how these terms relate to each other. This document aims to provide practical guidance for stakeholders regarding compliance with regulatory requirements labelled one way or another. With this goal in mind, it uses the umbrella term “explainability” and provides a non-exhaustive taxonomy and list of approaches that stakeholders can use to comply with regulatory requirements. While the overarching goal of explainability is to evaluate the trustworthiness of AI systems, at different stages of the AI system life cycle, diverse stakeholders can have more specific objectives in support of the goal. To illustrate this point, several examples are provided. For developers, the goal can be improving the safety
Scope: This document describes approaches and methods that can be used to achieve explainability objectives of stakeholders with regard to machine learning (ML) models and artificial intelligence (AI) systems’ behaviours, outputs and results. Stakeholders include but are not limited to, academia, industry, policy makers and end users. It provides guidance concerning the applicability of the described approaches and methods to the identified objectives throughout the AI system’s life cycle, as defined in ISO/IEC 22989.