Guidelines for Responsible AI
Bias and Fairness
Bias can enter through data, labels, feature choices, historical inequities, and feedback loops. Fairness work should happen across the lifecycle, not only after deployment.
Concepts
- Bias can enter through sampling, labels, features, historical inequities, or feedback loops.
- Fairness should be assessed across relevant groups and throughout the lifecycle.
- Class imbalance and proxy variables can create unfair outcomes even when protected attributes are removed.
Exam tips
- Evaluate model behavior across relevant groups, not only aggregate accuracy.
- Proxy variables can create unfair outcomes even if protected attributes are removed.
- SageMaker Clarify can help detect bias in data and models.