P15 · Responsible AI — Ethics, Safety & Governance

A screening model goes live. It is accurate on average, ships on time, passes its unit tests. Months later an audit finds it approved qualified applicants from one group far less often than another equally qualified group — not because anyone intended it, but because the training labels carried a historical bias the objective faithfully reproduced. No line of code was wrong. What was missing was the question this pillar exists to ask: is the system fair, private, and governable at all — and how would you know?

The first track is fairness and privacy made concrete. Where bias enters — data, labels, objective, deployment — walked end-to-end on one hiring cohort. Then the uncomfortable mathematics: demographic parity and equalized odds are both reasonable definitions of fairness, and you will prove, on a cohort you can total by hand, that when base rates differ you cannot satisfy both without a useless classifier. Privacy gets the same treatment: what "anonymised" data still leaks when you join it to a public list, k-anonymity, and the suppression rules that keep a published table from re-identifying a person. It closes on documentation — model cards and datasheets, the artefacts that make a system accountable.

The second track is governance and safety as a working skill, not a compliance checkbox. The regulatory landscape (EU AI Act risk tiers, accountability, human oversight) as a map a practitioner actually uses. LLM-specific failure modes — hallucination, prompt injection, misuse — through the safety lens. Safety evaluation and red-teaming, and the sober fact of what a passing eval does not prove. And when it goes wrong: incident thinking, walked through a real case end to end — detect, contain, communicate, remediate, prevent.