AI Ethics, Safety, and Governance
What this domain tests
Whether you can identify harm, measure fairness where metrics conflict, protect privacy, and place the controls - documentation, human oversight, red teaming - that regulators and reviewers expect before launch.
Core rules
- Fairness metrics conflict. Demographic parity, equal opportunity, and equalised odds generally cannot all hold when base rates differ; the choice is a policy decision that must be documented, not a technical default.
- Bias enters through historical labels, sampling, proxy variables, annotator bias, and deployment feedback loops that reinforce past decisions.
- Privacy: minimise and de-identify personal data, aggregate where possible, apply differential privacy for published statistics, and treat model memorisation as a real leakage path.
- Transparency: model cards, dataset documentation, disclosure that a user is interacting with AI, and audit trails.
- Safety: red teaming, jailbreak and prompt-injection testing, content filters, guardrails, and human-in-the-loop review for high-stakes decisions.
- Governance: classify risk (for example, prohibited, high-risk, limited-risk transparency obligations, minimal-risk tiers), run impact assessments, and name an accountable owner.
Worked example
A hiring model is 90% accurate overall, but the selection rate is 20% for group A and 45% for group B. Reasoning: apply the four-fifths rule - 20/45 is about 0.44, well under 0.8, so adverse impact is indicated despite high accuracy. Inspect proxy features such as postcode or school, reweight or remove them, reassess thresholds, and document the remaining trade-off with legal review.
Traps that cost marks
Presenting accuracy as evidence of fairness. Deleting protected attributes while leaving proxies intact. Treating explainability as optional in high-risk settings. Launching without human oversight or an appeals route. Scheduling the ethics review after deployment.