Responsible AI, Governance, and Best Practices
Why governance matters
Models influence decisions about people-credit, hiring, churn outreach. Responsible practice means documenting data provenance, intended use, limitations, and known risks, and ensuring the model is fair, explainable, and auditable.
Fairness and bias
Bias enters through biased historical data, proxy variables (e.g., ZIP code standing in for protected attributes), and sampling. Fairness metrics trade off against each other-equal accuracy across groups, equal opportunity, and calibration cannot all be maximized simultaneously. Questions often test whether you recognize a proxy variable or overstate a single fairness metric as universally correct.
Explainability and human oversight
Prefer interpretable models where stakes are high or regulation demands explanation. Provide per-decision explanations, allow human review of consequential outcomes, and avoid fully automated adverse decisions without appeal.
Privacy and documentation
Minimize personally identifiable information, apply access controls, and retain only what is needed. Maintain a model card or equivalent record: purpose, data, metrics, subgroup performance, owner, and review date.
Traps
Assuming a model is fair because protected attributes were excluded (proxies remain), or that high overall accuracy guarantees equitable subgroup performance. Another trap: treating governance as a one-time approval rather than ongoing monitoring.
Rule of thumb
If you cannot explain a decision to the affected person and to an auditor, the model is not ready. Check subgroup metrics, not just aggregate ones.