Bias Mitigation
Methods for identifying, reducing, and monitoring unfair bias in AI systems.
Did changing the training examples reduce missed requests?
Suppose a team adds independently reviewed examples of large-print requests phrased in an underrepresented language. It tests both model versions on the same held-out cases.
- Before
On 50 qualifying requests in language B, the original model flags 35. It flags 45 of 50 in language A.
- After
The revised model flags 42 of 50 in B and 44 of 50 in A. Observed recall rises from 70% to 84% in B and falls from 90% to 88% in A.
- Inspect the changes
The observed gap narrows from 20 to four percentage points. Inspect the newly found requests, the new misses and incorrect flags on nonqualifying requests.
Check that the improvement travels
The fixed comparison helps isolate what changed between model versions. Test on fresh branches and new phrasing before relying on the apparent gain. Ask staff and readers whether missed or unnecessary referrals carry different costs, then monitor those outcomes after deployment.
Why Bias Mitigation Is Ongoing
Bias mitigation is an ongoing governance practice, not a one-time optimization task. Real populations change. Business processes change. Data pipelines change. A model that looked acceptable in one setting may become unfair in another.
That is why good bias mitigation is tightly connected to monitoring, documentation, and accountability. The technical system matters, but so do the review process and the people responsible for acting on what they find.
Follow the idea
Model fairness compares behavior across groups for a specified task.
Model monitoring checks whether an observed improvement persists.
Related concepts (7)
Related Yenra articles
Source · Fairlearn guide to disaggregating evaluation metrics
Return to Bias Mitigation in the A–Z glossary