fairness (group vs individual)
/ FAIR-nis, groop vur-sus in-di-VIJ-oo-ul /
When we ask whether an algorithm is "fair," there are two very different things we might mean, and they often pull in opposite directions. Group fairness asks: are the outcomes balanced across whole groups — do men and women, or different ethnicities, get approved at similar rates? Individual fairness asks something narrower: are two people who are alike in every way that should matter treated alike, regardless of which group they belong to? The slogan for the second is "treat similar people similarly."
Here is the rub. Imagine a scholarship algorithm. To make the acceptance rate equal across two neighborhoods (group fairness), you might have to admit a slightly weaker applicant from one area over a slightly stronger one from another — which feels unfair to the two individuals being compared (a violation of individual fairness). Conversely, if you rank everyone purely by a "merit" score and the score itself was shaped by unequal schools, perfect individual fairness can lock in lopsided group outcomes. Researchers have proven that under realistic conditions you usually cannot satisfy every reasonable fairness definition at once. They genuinely conflict.
Why it matters: there is no single mathematical formula for "fair." Fairness is a value judgment about what kind of society we want, and the math only forces us to be explicit about the trade-offs. A team that says "our model is fair" without saying which definition it chose, and what it sacrificed to get there, hasn't answered the question — it has dodged it. The useful move is to name the harm you most want to avoid and pick the definition that guards against it.
Two loan applicants have identical incomes, debts, and credit histories — but live in different neighborhoods. Individual fairness says they must get the same decision. Group fairness might still flag a problem if, across thousands of such pairs, one neighborhood is approved far less often, revealing that the "identical" inputs were shaped by unequal histories.
The same case can look fair by one definition and unfair by the other — which is why you must choose, and say so.
There is a famous impossibility result: except in trivial cases, a model cannot be calibrated equally across groups AND have equal error rates across groups at the same time. So debates like the 2016 dispute over a US recidivism tool weren't about who did the math wrong — both sides were right, using different definitions of fair.