Overview
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 lifecycle‑based techniques to identify, mitigate and monitor unwanted bias in AI systems that use machine learning for classification and regression. Published as a first edition in 2024, the document is applicable to organizations of all sizes and provides context‑agnostic, practical mitigation approaches that can be applied across the AI system life cycle.
Key Topics
- AI system life cycle mapping: Mitigation actions aligned to inception, design & development, verification & validation, re‑evaluation/monitoring and disposal stages.
- Stakeholder and requirements analysis: Guidance on stakeholder identification, needs definition and acceptance criteria to reduce bias risks from the start.
- Data considerations: Emphasis on data sources, metadata sufficiency, data annotations, provenance and data quality measures (linked to ISO/IEC 5259‑4).
- Feature and model design: Feature representation, adjustments to training data, and handling of pre‑trained models to reduce bias in classification and regression tasks.
- Testing & validation: Static data testing, dynamic testing of models, and continuous validation strategies for operational monitoring.
- Risk integration: Integration with risk management and consumer vulnerability concepts to prioritize bias treatment.
- Techniques catalog: Algorithmic/training techniques and data techniques specifically targeted at treating unwanted bias.
- Distributed lifecycle handling: Considerations for systems developed or operated across multiple organizations or teams.
- Informative annexes: Life cycle process map and potential impacts of unwanted bias on different user groups.
Applications and Who Should Use It
ISO/IEC TS 12791 is intended for:
- AI/ML engineers and data scientists implementing classification or regression models who need concrete bias‑mitigation steps.
- Product managers and procurement teams defining acceptance criteria and supplier requirements to reduce bias risk early in projects.
- Risk, compliance and quality teams integrating bias treatment into organizational risk management and data quality frameworks.
- Regulators, auditors and standards professionals assessing whether an AI system follows lifecycle best practices for bias treatment.
- Organizations using or adapting pre‑trained models and needing guidance on validating and adjusting those models for fairness.
Practical uses include designing bias‑aware datasets, defining test plans for bias detection, integrating monitoring in production, and documenting decisions for transparency and governance.
Related Standards
- ISO/IEC 22989:2022 (AI concepts and terminology)
- ISO/IEC TR 24027 (types of bias - foundational reference)
- ISO/IEC 5259‑4:2024 (data quality for ML)
- ISO/IEC/IEEE 29119‑3:2021 (software test documentation)
- ISO/IEC 42001 (AI management system - related organizational processes)
Keywords: ISO/IEC TS 12791:2024, unwanted bias, bias mitigation, AI system life cycle, machine learning, classification, regression, data quality, model validation, pre‑trained models.
Открыть превью (PDF)Information technology — Artificial intelligence — Treatment of unwanted bias in classification and regression machine learning tasks