AI Ethics, Safety, Privacy and Societal Impact
What this domain tests
AI Ethics, Safety, Privacy and Societal Impact tests bias, data protection, safety, employment effects and social consequences. Answers must connect ethical principles to concrete design choices, not just list slogans.
Core rules and concepts
- Bias can enter through historical data, sampling, labels, measurement or deployment. Fairness requires auditing outcomes across groups, not only checking code.
- Privacy principles include consent, purpose limitation, data minimisation and storage limitation. Anonymisation and pseudonymisation reduce risk, but re-identification is still possible. India's DPDP Act 2023 and the EU GDPR are relevant legal examples.
- Safety covers robustness, human oversight, misuse prevention and accountability. Deepfakes and automated decisions can cause harm at scale.
- Employment effects include displacement of routine tasks and creation of new roles. Benefits and harms are unevenly distributed, creating a digital divide.
Worked example
A hiring model trained on past hires favours men because past managers hired mostly men. The model achieves high accuracy but low fairness: it may learn gender proxies such as college or hobby. Reasoning: the label 'successful hire' encodes historical bias. Remedies include auditing disparate impact, removing proxy features, reweighting data and keeping human review.
Traps that cost marks
Bias is not only a data problem; model design and deployment also create it. Anonymisation is not an absolute guarantee. Ethics is broader than legal compliance. Do not claim AI is objective because it uses numbers. Do not ignore human oversight in high-stakes decisions.