Overview
CEN/CLC ISO/IEC/TS 12791:2024 - Information technology - Artificial intelligence - Treatment of unwanted bias in classification and regression machine learning tasks - is a Technical Specification that describes how to identify and treat unwanted bias in AI systems that use machine learning for classification and regression. Applicable to organizations of all sizes and sectors, the document provides bias mitigation techniques mapped to the AI system life cycle, aligned with ISO/IEC TR 24027 and life‑cycle definitions in ISO/IEC 22989 and ISO/IEC 5338.
Key Topics
- AI lifecycle integration: Guidance for treating bias at key stages - inception, design and development, verification and validation, re‑evaluation/monitoring, operations and disposal.
- Stakeholder processes: Stakeholder identification, requirements definition and procurement considerations to reduce bias risks early.
- Data practices: Metadata sufficiency, data annotation, data source evaluation and methods for adjusting data to mitigate bias.
- Algorithmic and training techniques: High‑level approaches for algorithmic mitigation, including considerations for pre‑trained models.
- Testing and validation: Static testing of development data, dynamic testing of models, acceptance criteria, and continuous validation during operation.
- Risk management: Integration with organizational risk processes and handling bias in distributed AI system life cycles.
- Supporting resources: Informative annexes (life‑cycle process map, potential impacts on specific user types) and bibliographic references.
Practical Applications
Who uses ISO/IEC TS 12791 and how it’s applied:
- Data scientists & ML engineers: Implement data techniques and algorithmic mitigations during model training and evaluation to improve fairness.
- Product managers & procurement officers: Define acceptance criteria and supplier requirements to reduce bias risk when acquiring models or datasets.
- Compliance, risk & governance teams: Integrate bias treatment into risk frameworks and continuous monitoring programs.
- DevOps/ML Ops teams: Apply dynamic testing, monitoring and re‑evaluation in production to detect and treat emergent bias.
- Regulators and auditors: Use the standard as a reference for assessments of bias treatment practices in classification and regression systems.
Related Standards
- ISO/IEC TR 24027 (types of bias and terminology)
- ISO/IEC 22989 and ISO/IEC 5338 (AI system life cycle)
- ISO/IEC 42001 (AI management systems) - note: TS 12791 does not cover AI management system processes contained in 42001.
This specification is a practical, life‑cycle oriented resource for implementing bias mitigation, AI fairness, and robust machine learning governance in classification and regression applications.