ISO/IEC 12792:2025 (E) - Information technology � Artificial intelligence (AI) � Transparency taxonomy of AI systems

0.1 General

This document specifies a taxonomy of information elements to assist AI stakeholders with identifying and addressing the needs for transparency of AI systems. The taxonomy helps organize disclosures across the AI lifecycle, supporting explainability, accountability, and trustworthiness.

It is applicable to any kind of organization and application involving an AI system, regardless of size or sector.

Example: A developer of a generative AI tool for content creation can use this taxonomy to document data sources, model behaviors, and intended uses for users and regulators.

0.2 Purpose of the Transparency Taxonomy

The taxonomy structures transparency information into layers (Context, System, Model, Dataset) to address varying stakeholder needs, from developers to end-users and auditors. It promotes consistent terminology and supports compliance with regulations such as the EU AI Act (Article 13 on transparency obligations).

Example: In a credit scoring AI system, the taxonomy ensures disclosure of decision-making logic to affected individuals, addressing fairness and non-discrimination concerns.

This document defines a taxonomy of information elements for transparency of AI systems. It covers semantics of elements and their relevance to stakeholder objectives across the AI lifecycle.

It does not specify implementation methods or detailed disclosure formats but provides a structured framework for transparency documentation.

Example: For an AI used in medical diagnostics, the scope includes transparency items related to training data, model performance, and deployment context.

The following documents are referred to in the text in such a way that some or all of their content constitutes requirements of this document:

- ISO/IEC 22989, Information technology � Artificial intelligence � Artificial intelligence concepts and terminology

- ISO/IEC 42001, Artificial intelligence � Management system

- ISO/IEC 23894, Information technology � Artificial intelligence � Guidance on risk management

Example: Use ISO/IEC 22989 for consistent terminology when applying this transparency taxonomy.

Key terms include:

- Transparency: Openness about decisions and activities of an AI system that affect stakeholders, communicated in a clear, accurate, and complete manner.

- AI System: Engineered or machine-based system that generates outputs influencing real or virtual environments based on given inputs.

- Taxonomy: Structured classification of information elements for transparency.

Example: "Residual Risk" in transparency contexts refers to uncertainties remaining after disclosures about model limitations.

4.1 Concept of Transparency

Transparency in AI involves providing relevant information about the system's purpose, functioning, data, and limitations to enable informed use and oversight.

Example: Disclosing that a facial recognition system performs differently across demographic groups promotes informed deployment decisions.

4.2 Benefits of Structured Transparency

Facilitates trust, regulatory compliance, risk management, and collaboration among stakeholders.

Example: Regulators can better assess conformity; users can understand limitations; developers can improve systems iteratively.

5.1 General

This clause provides an overview of the taxonomy and its application to AI systems.

5.2 Transparency Layers

The taxonomy is organized into four primary layers: Context-level, System-level, Model-level, and Dataset-level.

Example: Context-level covers intended use and stakeholders; Dataset-level details training data characteristics.

6.1 Identification of Stakeholders

Identify relevant parties such as developers, deployers, users, affected individuals, auditors, and regulators.

Example: End-users may need simple explanations, while auditors require detailed technical data.

6.2 Transparency Objectives

Objectives include enabling understanding, accountability, fairness assessment, and risk mitigation.

6.3 Tailoring Disclosures

Customize information depth and format based on stakeholder needs and risk level.

7.1 General Context Information

Information about the broader environment, objectives, and deployment scenario of the AI system.

7.2 Intended Purpose and Use Cases

Describe what the AI is designed for, limitations, and prohibited uses.

Example: "This AI assists in medical image analysis for radiologists but is not a replacement for professional diagnosis."

7.3 Organizational and Regulatory Context

Details on developer organization, compliance references, and applicable standards.

8.1 System Architecture and Components

High-level description of how the AI system operates as a whole.

8.2 Interaction with Environment

Inputs, outputs, human-AI interaction, and integration points.

8.3 Performance and Limitations

Metrics, known biases, robustness, and safety considerations.

9.1 Model Type and Methodology

Details on algorithms, hyperparameters, and design choices.

9.2 Training and Evaluation

Information on training procedures, validation methods, and performance benchmarks.

9.3 Interpretability and Explainability

Techniques used to make model decisions understandable.

10.1 Dataset Description

Characteristics, size, sources, and composition of datasets used.

10.2 Data Collection and Processing

Methods for acquisition, cleaning, labeling, and augmentation.

10.3 Quality and Bias Considerations

Assessments of representativeness, completeness, and potential biases.

Example: Document demographic distribution in training data to support fairness analysis.