Fairness means treating relevant people or groups equitably in an AI system’s behavior and outcomes. It is a product, governance, and trust concern, not merely a model-performance metric. section 1.5 section 11.0
Build fairness into the AI lifecycle:
- Inspect training data for historical discrimination, missing representation, and skewed patterns. section 11.3.4
- Run bias detection and predeployment fairness testing, then apply structured de-biasing techniques where needed. section 11.3.4
- Flag predictions affected by limited or unbalanced data, especially for less active regions or groups. section 12.2.4
- Monitor deployed systems, refresh data, explain decisions, and communicate limitations clearly to affected stakeholders. section 11.3.4 section 1.5
Amazon hiring algorithm
A hiring tool trained on historical decisions penalized female candidates, reproducing past gender bias. The failure illustrates why representative data, fairness audits, and predeployment testing matter. section 11.3.4
