Why does removing race and gender columns from the training data fail to make a model fair?

## The claim being tested "We don't collect race, so the model can't be racist." This position has a name in the fairness literature — **fairness through unawareness** — and it is the first thing an interviewer wants you to be able to take apart. A **protected attribute** is a characteristic that law or policy forbids you from using as a basis for a decision: race, sex, age, disability, religion, national origin, and in some jurisdictions pregnancy or marital status. Fairness through unawareness is the policy of simply not putting that column in the feature matrix. ## Why it fails: redundant encoding A feature is a **proxy** for a protected attribute when it carries information about that attribute. A model does not need the attribute itself; it needs anything correlated with it plus a label whose historical pattern differs by group. Three mechanisms produce proxies: 1. **Geography.** Residence is strongly segregated in many countries, so ZIP code, census tract, neighbourhood, or even distance-to-office is a partial encoding of race and often of income. An auto-insurance pricing model built on ZIP-level features can produce systematically different prices by race with no race column anywhere in its training table — the mechanism people call digital redlining. 2. **Institutions and affiliations.** A women's college, a women's chess club, a religious school, a national-origin-linked professional body: each is close to a deterministic marker of the attribute. A resume screener trained on which past applicants were hired can learn to downweight exactly those tokens if the historical hiring pattern did. 3. **Behaviour.** Purchase categories, browsing, device type, name spelling, and language of use all shift by group. The subtle version is **joint reconstruction**. No single feature has to look suspicious. Several features that are individually weakly correlated with the attribute can, in combination, reconstruct it with high accuracy — this is ordinary supervised learning applied to the attribute rather than to your label. That is why a per-feature correlation screen is not a sufficient audit. ## How to actually check The direct test is to treat the protected attribute as a target: hold the attribute out of the model's feature set, then fit a separate model that tries to *predict the attribute* from the remaining features. If that model beats the base rate substantially — say it separates the groups far better than chance — your feature set encodes the attribute, and any downstream model is free to use that encoding. Supporting checks: compare each feature's distribution across groups, and look at how much a feature's predictive contribution changes when you condition on group. Running this test requires you to *have* the attribute. That is the second lesson of unawareness: not collecting the attribute at all makes the audit impossible while leaving the mechanism intact. The workable arrangement is to collect it, exclude it from the feature set, and store it as an audit-only column with restricted access, used for measurement and nothing else. ## What to do once you find a proxy Dropping the proxy is not an automatic win. Three considerations: - **The disparity may survive.** Remove ZIP and the remaining features often re-encode much of it. You measure the disparity before and after rather than assuming. - **You may destroy legitimate signal.** ZIP code genuinely predicts theft and collision risk for reasons that are not race. Removing it costs accuracy for everyone, including members of the group you meant to protect. - **The question is what the feature is doing.** A feature that predicts the outcome only through group membership is a pure proxy. A feature with a defensible causal path to the outcome is a business-necessity argument you can make out loud. Interviewers care that you can tell these apart and can say which one you think you have. ## What good sounds like in an interview Name the failure mode ("fairness through unawareness"), give one concrete proxy mechanism, make the joint-reconstruction point so it is clear you know a single-feature correlation screen is not enough, and finish on measurement: the reason to keep the attribute is to compute outcomes by group, not to feed it to the model. That last move is what separates a candidate who has read about fairness from one who has had to run an audit.