Saudi Arabia SDAIA AI Ethics Principles (Version 2.0, 2025) 1.BV: Fairness in Build and Validate: fairness-conscious feature choices, champion model within fairness thresholds, causal feature selection and safeguards against discriminatory deployment

What does 1.BV require?

Engineers and designers keep fairness in view throughout build, test and implementation, aware that how features are grouped or split, kept or dropped, and judgements about the reliability and security of the feature set, can have serious consequences for vulnerable or protected groups. The champion model is selected with its fairness metrics within the defined thresholds for sensitive features, the approach to optimising fairness against performance is set out, and a failing assessment is justified before proceeding; feature selection is causality-based and verified with business owners and non-technical teams; and mechanisms are set up to prevent harmful and discriminatory results from automated decision support at deployment.