Responsible AI
Fairness & Bias
Bias can enter through data, labels, objectives, deployment context, and feedback loops. Fairness is not one universal metric; it is a product, policy, and measurement decision shaped by the harm model.
- Historical data can encode historical unfairness
- Labels can be biased proxies
- Removing sensitive attributes is not enough
- Fairness metrics can conflict
- Slice evaluation is the minimum bar
- Fairness requires governance, not just model tuning
| Stage | How bias enters |
|---|---|
| Data collection | Some groups are underrepresented or measured differently |
| Labels | Human/system labels encode prior decisions |
| Features | Proxy variables reconstruct sensitive attributes |
| Objective | Metric rewards behavior that harms a group |
| Deployment | Users interact with the system differently |
| Feedback loop | Model decisions shape future data |