Bias and Fairness Control
Measuring and constraining disparate outcomes across groups, where the definition of fairness must be chosen deliberately because the definitions are mutually incompatible.
The uncomfortable technical fact at the centre of this: several reasonable definitions of fairness cannot be satisfied simultaneously except in degenerate cases. Demographic parity requires equal positive rates across groups. Equalised odds requires equal error rates. Calibration requires that a given score mean the same thing for everyone. Where base rates differ between groups, these are provably incompatible.
So "make it fair" is not an implementable requirement. Someone must decide which definition applies, and that is a decision for the business, legal and the affected domain — with the reasoning recorded, because it will be questioned.
The sources of disparity are worth separating because they need different remedies. Historical bias in training data reflects past discrimination and is reproduced faithfully by a model that learns it. Representation bias comes from under-sampled groups, producing worse accuracy where there was less data. Proxy variables encode a protected attribute indirectly — postcode for ethnicity, for instance — so removing the protected field alone achieves nothing.
The controls: disaggregated evaluation across relevant groups as a standing requirement rather than a launch check, monitoring for drift in disparity after deployment, an appeal path for individuals, and documentation of the fairness definition chosen. Note that measuring disparity requires holding data on the protected attribute, which sits in tension with minimisation — a genuine conflict that needs a deliberate, documented resolution rather than being resolved by accident.