Fairness & Privacy
A track of P15 · Responsible AI — Ethics, Safety & Governance.
Where bias enters, fairness metrics that provably conflict, what aggregation cannot hide, and the documentation that keeps a model accountable.
Fairness and privacy are where good intentions meet hard constraints. This track refuses to leave either at the level of a slogan: both are worked on a tiny cohort and a tiny dataset you can check by hand, so the claims are demonstrated, not asserted.
It opens with where bias actually enters a model — data, labels, objective, deployment — traced through one hiring cohort so the abstraction becomes a story you can follow. Then the mathematics that surprises people: demographic parity (equal approval rates) and equalized odds (equal error rates) are both defensible definitions of fairness, and you will prove — on the cohort, with a threshold sweep — that when two groups have different base rates you cannot satisfy both at once without throwing away a useful model. Privacy gets the same concrete treatment: what "anonymised" health rows still leak when an attacker joins them to a public voter list, the k-anonymity that measures the exposure, and the small-cell suppression (with its complementary-suppression trap) that a published table needs. The track closes on the documentation — model cards and datasheets — that turns all of this into something a reviewer can hold a system to.