AI Ethics, Bias, and Responsible AI
Responsible AI rests on principles: fairness, transparency, accountability, safety, privacy, and human oversight. Know the difference between ethics (what ought to be done) and compliance (what is legally required); a system can be legal and still unethical.
Bias enters through historical data (reflecting past discrimination), sampling (underrepresented groups), labeling (annotator bias), and proxy variables (e.g., ZIP code standing in for race). Distinguish harm types: allocative (resources denied), representational (stereotyping), and quality-of-service (worse performance for a subgroup). Fairness metrics conflict - demographic parity (equal positive rates), equalized odds (equal error rates), and individual fairness cannot generally all hold at once; choosing one is a values decision.
Explainability matters for recourse and audit; some models are inherently interpretable (linear, decision trees) while others require post-hoc methods (feature importance, SHAP). Transparency means disclosing that AI is in use and how decisions are made.
Traps: assuming removing a protected attribute removes bias (proxies remain), assuming a model is fair because overall accuracy is high (subgroup performance may differ), or treating explainability as identical to fairness.
Worked example: a hiring model with 90% overall accuracy may show 85% recall for group A and 60% for group B. Before deployment, compute metrics per subgroup, inspect proxy features, and decide which fairness criterion the business and legal context require.
Rule of thumb: evaluate performance and harm by subgroup, not just in aggregate, and document the tradeoff you chose.