Bias and fairness

Learned from the past.

How it enters

Training data records past decisions, historical inequalities and unrepresentative samples. A model reproduces those patterns and applies them faster and at larger scale.

Removing an attribute rarely helps, because proxies remain.
Under-representation shows up as worse accuracy for smaller groups.

Fairness is plural

Equal error rates, equal outcomes and equal treatment are mathematically incompatible in most realistic cases. Someone must choose which definition applies, and that choice should be stated.