First edition 2025-09 | Version: Illustration
This technical specification provides guidance on explainability and interpretability for AI systems, focusing on objectives and approaches to make ML models and AI decisions understandable.
It is distributed for review and comment, subject to change, and not to be referred to as an International Standard until published as such.
Example: Organizations can use this to improve trust in AI by explaining how a loan approval model makes decisions.
This document addresses the need for explainability in AI to build trust, ensure accountability, and comply with regulations like the EU AI Act.
It covers concepts, objectives, and methods for making AI decisions interpretable.
Example: In healthcare AI, explainability helps doctors understand why a diagnosis was made, improving acceptance.
This technical specification provides objectives and approaches for explainability and interpretability of ML models and AI systems.
It applies to organizations developing or using AI where understanding decisions is crucial.
Example: Used in financial AI to explain credit denials, aiding compliance with fairness laws.
No normative references are cited in this document.
Example: Informative references include ISO/IEC 22989 for AI concepts.
Key terms:
- Explainability: Extent to which AI decisions can be understood by humans.
- Interpretability: Ability to understand the model's internal workings.
- AI System: Machine-based system with autonomy.
Example: "Black-box Model" is one where internal logic is not interpretable, like deep neural networks.
Abbreviations:
- AI: Artificial Intelligence
- ML: Machine Learning
- XAI: Explainable Artificial Intelligence
Example: Use "XAI" to refer to systems designed for explainability.
This clause discusses human oversight in AI, including monitoring and control to ensure safe operation.
It covers approaches like pre-task, in-task, and post-task oversight.
Example: In autonomous driving AI, human oversight might involve monitoring dashboards for intervention.
Provides examples of oversight, like interpretability tools and human intervention points.
Example: Use LIME for local explanations in ML models.
References to related standards like ISO/IEC 22989 and publications on AI explainability.
Example: Cite "Nudging and Boosting: Steering or Empowering Good Decisions" for decision support.