AI Ethics, Bias, Privacy and Societal Impact
What the exam tests
This domain asks you to apply responsible-AI principles to concrete cases, identify sources of bias, explain data-protection obligations, and discuss how automation changes work rather than repeating slogans.
Core concepts
- Responsible AI principles: fairness, transparency and explainability, accountability, privacy, safety and robustness, and human oversight.
- Bias sources: historical bias (past discrimination baked into labels), sampling bias (unrepresentative data), measurement bias (a proxy that measures the wrong thing), and label bias (inconsistent human labelling).
- Privacy: collect only what is needed (data minimisation), obtain informed consent, state the purpose and stick to it, and secure the data. Anonymisation removes identifying links; pseudonymisation replaces identifiers with codes but can be reversed with the key.
- India's Digital Personal Data Protection Act, 2023 governs personal data: it requires consent or a legitimate use, purpose limitation, and gives data principals rights such as access and correction.
- Societal impact: automation displaces some routine tasks while creating data, annotation, and oversight roles; it can also widen gaps if access to AI tools and skills is unequal.
Worked example
A company trains a resume-screening model on ten years of past hiring decisions, 85% of which favoured male candidates. Reasoning: the model learns the historical pattern, so it penalises resumes associated with women. This is historical bias, not a coding error. Fixes: audit the training data, remove or reweight biased features, measure selection rates across groups, add human review, and document limitations. Removing gender alone is insufficient because proxies such as college or hobby names can still encode it.
Traps that cost marks
- Defining bias as only a technical bug; most bias enters through data and past decisions.
- Saying anonymisation and pseudonymisation are identical.
- Treating consent as a one-time checkbox rather than specific, informed, and withdrawable.
- Assuming removing a protected attribute guarantees fairness.
- Claiming AI will simply 'destroy' or 'create' all jobs; the exam expects balanced analysis of task-level change.