AI Ethics and Societal Impact
Domain 2 - AI Ethics and Societal Impact. The exam tests whether you can name specific ethical principles, apply them to a concrete system, and distinguish legal compliance from ethical acceptability.
Core rules and concepts
- Six recurring principles: fairness, transparency and explainability, accountability, privacy, safety and robustness, and human oversight.
- Bias enters through data, labels, features and deployment. Historical bias becomes label bias; proxy variables (a postcode standing in for caste, religion or income) produce indirect discrimination even when the sensitive attribute is removed.
- Privacy: purpose limitation, data minimisation, informed consent, and correction or erasure rights. India's Digital Personal Data Protection Act, 2023 requires notice plus consent for processing personal data, with limited exemptions.
- Accountability: a named human or organisation must answer for harm; the algorithm decided is not a defence. High-stakes uses - health, hiring, credit, policing, education - need explanation and an appeal route.
- Regulation: the EU AI Act (in force 2024) sorts systems into unacceptable risk (banned), high risk (heavy obligations), limited risk (transparency duties) and minimal risk.
- Societal impact: automation of routine work, reshaping rather than simple elimination of jobs, deepfakes and misinformation, filter bubbles, and unequal access (the digital divide). Asimov's Three Laws are fiction, not engineering requirements.
Worked example
A company trains a CV-screening model on ten years of hires, most of them men, and the model downgrades CVs mentioning a women's chess club. Diagnosis: label bias plus proxy features, because the historical outcome encodes past discrimination. Remedy: audit features for proxies, rebalance or reweight the training data, test outcomes across groups, keep a human reviewer, and publish an appeal route. Removing the word women alone fails, because correlated proxies survive.
Traps that cost marks
- Treating legal and ethical as identical: something can be lawful and still unfair.
- Saying the data is objective, therefore the model is fair.
- Confusing explainability with accuracy; a very accurate model can be unexplainable.
- Assuming anonymisation alone protects privacy, when linkage can re-identify people.