AI Ethics, Bias, and Responsible AI
Sources of bias
Bias enters through historical data, sampling, labeling, feature selection, and deployment context. Distinguish representation bias (groups underrepresented in data), measurement bias (a proxy measures the wrong thing), and label bias (annotator or process skew). Understand that removing a protected attribute does not remove bias, because correlated proxies remain.
Fairness concepts
Know that fairness definitions conflict: demographic parity (equal positive rates), equalized odds (equal error rates across groups), and individual fairness (similar treatment for similar individuals). Satisfying all simultaneously is often impossible, so the choice is a policy decision tied to context.
Responsible AI principles
Typical principles include fairness, transparency and explainability, accountability, privacy, safety, and human oversight. Explainability techniques split into inherently interpretable models and post-hoc methods (feature importance, local explanations). Know that explanations can be approximate and should not be over-trusted.
Governance in practice
Responsible practice includes impact assessments, diverse review, documentation (model cards, datasheets), red-teaming, and monitoring for disparate performance after deployment.
Worked example
A hiring model excludes gender but uses zip code and career-gap features that correlate with gender, producing disparate outcomes. The fix is measuring outcomes by group, auditing proxies, and choosing a fairness criterion with stakeholders, not simply deleting a column.
Rule of thumb
Fairness is measured on outcomes across groups, and the acceptable tradeoff must be defined before evaluation.