Fairness means equitable outcomes across relevant groups—defined contextually, not one formula. Accountability assigns clear ownership for model behavior, incidents, and remediation.
Fairness definitions (preview)
- Demographic parity — similar approval rates across groups
- Equal opportunity — similar true positive rates
- Calibration — predicted risk matches observed risk within groups
Definitions can conflict—teams must choose based on harm analysis and law.
Accountability structures
- Named model owner and on-call rotation
- Incident runbooks for harmful outputs
- Audit logs: who deployed which version when
- Post-incident reviews without blame-shifting to "the algorithm"
Disparate impact awareness
Even neutral features (zip code) can proxy protected attributes. Document feature rationale and monitor sliced metrics ongoing—not only at launch.
Important interview questions and answers
- Q: Fairness metrics conflict?
A: Satisfying one mathematical fairness definition may violate another—requires policy choice. - Q: Accountable to whom?
A: Users, regulators, internal risk committees, and affected communities.
Self-check
- Name two fairness definitions.
- What should a model owner document?
Pitfall: Picking a fairness formula without policy input—definitions conflict by design.
Interview prep
- Fairness metrics conflict?
- Demographic parity, equal opportunity, and calibration cannot all hold simultaneously in general.
- Model owner role?
- Accountable for incidents, documentation, and remediation—not blameless "algorithm."