Module 6 — AI Ethics, Bias & Governance

AI bias is not a theoretical concern — it has caused real harm. Amazon's hiring AI discriminated against women because it was trained on historical hiring data that favoured men. COMPAS, a criminal justice AI, showed racial bias in recidivism predictions. Facial recognition systems have significantly higher error rates for people with darker skin.

Bias enters AI systems through multiple pathways:

Historical bias: Training data reflects past discrimination. If historical lending data shows fewer loans to minorities, an AI trained on this data perpetuates the pattern.

Representation bias: Training data doesn't adequately represent all groups. If a facial recognition system is primarily trained on lighter-skinned faces, it performs poorly on darker-skinned faces.

Measurement bias: The features used to train the model are proxies for something else. Using "zip code" as a feature may effectively encode race due to residential segregation.

Aggregation bias: One-size-fits-all models may not work equally well for different subgroups.

Fairness in AI is complex because different definitions of fairness can be mathematically incompatible. Demographic parity (equal prediction rates across groups), equalised odds (equal error rates), and individual fairness (similar people treated similarly) cannot all be satisfied simultaneously — hard trade-offs are necessary.

The EU AI Act (2024) is the world's first comprehensive AI regulation. It takes a risk-based approach:

Unacceptable risk (banned): Social scoring by governments, real-time biometric surveillance in public spaces (with exceptions), manipulation of vulnerable groups.

High risk (strictly regulated): AI in critical infrastructure, education, employment, essential services, law enforcement, immigration. These require risk assessment, human oversight, transparency, and documentation.

Limited risk (transparency obligations): Chatbots must disclose they're AI; deepfakes must be labelled.

Minimal risk (no restrictions): AI-enabled games, spam filters, most consumer applications.

Responsible AI frameworks provide principles for building trustworthy AI:

Transparency: People should understand how AI decisions are made. Accountability: There must be clear responsibility for AI outcomes. Fairness: AI should not discriminate or perpetuate bias. Privacy: Data should be collected and used ethically. Safety: AI systems must be reliable and secure. Human oversight: Humans should remain in control of consequential decisions.

Organisations like Google, Microsoft, and the OECD have published AI principles. The challenge is translating principles into practice — embedding ethics into the entire AI lifecycle from design through deployment and monitoring.