Governance, Bias, and Emerging Technology Ethics
What the exam tests
Governance, Bias, and Emerging Technology Ethics tests how organizations make ethical decisions repeatable and how they detect and mitigate bias in new systems. Questions cover governance structures, documentation, impact assessment, fairness metrics, human oversight, and emerging-technology risks.
Core rules and concepts
- Governance needs authority, not just an ethics statement: accountable owners, policies, review gates, escalation, training, audit, and remedies.
- Impact assessments should be proportionate and documented: purpose, necessity, stakeholders, harms, mitigations, monitoring, and go/no-go criteria.
- Bias can enter through history, sampling, measurement, labels, proxies, deployment context, and feedback loops. Removing a protected attribute does not remove bias because proxies remain.
- Fairness metrics encode different values: demographic parity equalizes outcomes; equalized odds equalizes error rates; calibration aligns predicted probabilities. They can conflict, so choose with justification.
- Emerging systems such as generative AI, biometrics, and autonomous tools add risks: hallucination, misuse, opacity, surveillance, environmental cost, and automation bias. Human oversight must be meaningful, resourced, and able to stop the system.
- Vendors share responsibility, but the deployer remains accountable for fit, notice, monitoring, and redress.
Worked example
A resume screener is trained on ten years of hiring data that favored one demographic. History bias enters the labels; proxy bias enters features such as school or career gaps. The team should define the decision context, test error rates by group, choose a fairness target with legal and ethical rationale, remove or transform proxies, add human review for adverse decisions, and monitor outcomes after deployment. If no threshold can be met, do not deploy. Documentation should record the trade-off and the appeal route.
Traps that cost marks
Treating one fairness metric as universally correct. Believing bias is only a data problem. Treating an ethics board as a substitute for controls. Assuming vendor contracts transfer all accountability. Adding 'human in the loop' without authority or time to review. Ignoring feedback loops and post-deployment drift.