Responsible AI, Ethics, and Governance
What to know
Expect fairness definitions and their incompatibility (demographic parity, equalized odds, predictive parity cannot all hold simultaneously except in degenerate cases), bias sources (historical, sampling, label, measurement), transparency and explainability (feature importance, SHAP, LIME, inherently interpretable models), privacy (anonymization limits, differential privacy, k-anonymity), consent and data minimization, accountability, and regulatory frameworks (risk-based governance, documentation such as model cards and datasheets). Understand human-in-the-loop and contestability.
Traps
- Assuming removing a protected attribute eliminates bias - proxies remain.
- Treating fairness as a single metric rather than a choice with tradeoffs.
- Believing anonymization is permanent; re-identification via linkage is common.
- Claiming post-hoc explanations are causal.
- Ignoring feedback loops where predictions shape future training data.
Rule of thumb
State the fairness criterion explicitly, measure disparities across groups on validation data, document limitations, and provide recourse. Fairness is a design decision, not a default.
Worked example
A hiring model has equal accuracy across groups but selection rates of 30% vs. 15%. Demographic parity fails (ratio 0.5, below the common 0.8 four-fifths rule) even though accuracy looks fine - a classic exam scenario testing metric choice.