A team removes the ethnicity field from a credit model's training data and tells the risk committee the model can no longer discriminate. What is the most accurate response?

The mechanism

A model with a hundred features does not need a column labelled ethnicity, because the information is redundantly encoded across the features that remain. Postcode carries residential segregation. Employer, education, device type, name-derived features and transaction patterns each carry a fraction. A model optimising for repayment will use whatever combination predicts, and the combination that predicts can track the attribute closely.

The practical test is one afternoon of work: train a classifier to predict the protected attribute from the remaining features. If it does substantially better than chance — and on consumer data it usually does — then the attribute is still present in the input, and so is the pathway for an outcome gap.

This is the position known as fairness through unawareness. It has one real property: it removes the attribute as an explicit input, which matters legally in places where using it is prohibited. It provides no assurance about outcomes.

Why the other options fail

"The model is now fair because it cannot see the attribute." This confuses the input with the outcome. Fairness claims are about the distribution of decisions, and the decision distribution is unchanged by deleting a column the model can reconstruct.

"Fairer but less accurate." This assumes an accuracy-fairness trade that the change does not actually make. Dropping one redundant column usually costs almost nothing in accuracy and buys almost nothing in outcomes. It is the familiar trade-off invoked in the wrong place.

"Removing the field is required before testing for disparate outcomes." The opposite. You need the attribute to measure the gap, even when you may not use it to decide. The correct architecture keeps it out of the feature store and in a separate measurement store with restricted access — removing it from both is the common and damaging version of this mistake, because it makes the organisation permanently unable to answer the question.

What actually works

• Measure outcomes by group: approval rates, score distributions, error rates by group, with confidence intervals.
• Keep the attribute available for measurement only, in a separate store with its own access control, or infer a documented probabilistic proxy where collection is not permitted.
• Choose the fairness definition explicitly and write down who chose it. Equal approval rates and equal error rates cannot both hold when base rates differ.
• Re-measure after every retrain, because the proxies shift when the data does.

When removing the field is fine and when not

When the law forbids using it, remove it from the features — that is a legitimate and sufficient reason, and it is what most regulated lenders do. Just do not let the removal be reported as a fairness control. In the narrow case of a model with a handful of inputs that genuinely carry no group information — five clinical measurements, say — unawareness plus outcome monitoring is a reasonable position. The monitoring is the part doing the work in both cases.