Fairness

Fairness in AI means that a system's outputs don't produce unjustified disparities across groups defined by race, gender, age, disability, or other protected characteristics. It requires deliberate design and ongoing evaluation — not because organizations intend to discriminate, but because the data, objectives, and historical patterns that models learn from often encode existing disparities. There is also no single mathematical definition of fairness; different metrics (equal selection rates, equal error rates, calibrated probabilities) can be mutually incompatible, which means fairness criteria need to be chosen and documented, not assumed.

A model can be statistically accurate on average while producing outcomes that are systematically worse for specific groups — and the aggregate metric will never reveal it. Fairness and accuracy are not the same thing. In hiring, lending, pricing, and benefits decisions, this matters both ethically and legally: disparate impact under fair employment or lending laws can trigger regulatory enforcement regardless of whether discrimination was intended. Fairness is also not a one-time check — it requires monitoring after deployment, because models drift, populations shift, and yesterday's compliant system can become tomorrow's liability.