Dropping sensitive attributes does not equal equal outcomes

What it is

Fairness through unawareness is the intuitive move: delete gender, race, and age from the feature table, and since the model cannot "see" them, its output cannot discriminate. The reasoning fails. Removing the sensitive column deletes only the explicit label; the model can still recover the attribute's information from combinations of remaining features, and group disparities in outcomes persist — now without an audit handle. The right question is never "are sensitive attributes in the inputs" but "how are outputs distributed across populations."

Why it happens

Sensitive attributes are informationally redundant across feature spaces: postal codes and neighborhood composition, purchase categories and income, phrasing and language background, device model and economic tier, names and gender or cultural origin — any combination suffices for the model to reconstruct the deleted attribute to usable accuracy. The model does not need to "know" race to execute race-correlated patterns: it learns behavioral and situational features strongly correlated with the attribute, which is operationally equivalent to using it. Worse, deletion drives discrimination underground: with the column present one could directly test "does the attribute change output, holding others constant"; after deletion even that test is impossible — auditors cannot see the attribute either, so disparate impact becomes less likely to be detected. Legal practice has accordingly clarified that "we didn't use sensitive attributes" does not discharge disparate-impact responsibility.

Studying it

Ablation audits are standard: compare group output distributions of models with and without sensitive attributes; if the gap is essentially unchanged, unawareness has failed. Reconstruction experiments verify from the other side: train a predictor of the sensitive attribute using only allowed features — accuracy far above chance proves the information survives. Variables include feature-set richness (more features, more redundancy), group base rates, and geographic granularity. Methodological cautions: reconstruction experiments carry an ethical edge — proving reconstructability also manufactures inference capability for sensitive attributes, so experimental data must be isolated and destroyed; audit labels come from self-report or proxies, and misclassification rates must be reported alongside conclusions.

Where it stops holding

"Deletion fails" does not mean "keep sensitive attributes always": under specific constraints (sparse features weakly correlated with the attribute, simple models), deletion can genuinely weaken direct effects; and some jurisdictions do forbid processing sensitive attributes for particular decisions. There compliance requires deletion — but deletion must then shift fairness supervision to output-distribution audits: input-side prohibitions and output-side accountability run in parallel. The sole criterion of "no disparity" remains the group distribution of outcomes, not the cleanliness of features.

Applying it

• Redirect fairness audits from input lists to output distributions: whatever is in the feature table, report pass rates, error rates, and score distributions stratified by population.
• Retain (where lawful) or separately collect sensitive attributes for evaluation and monitoring, with evaluation data physically separated from training data and routed only into the audit pipeline.
• Run a reconstruction experiment before launching high-stakes decisions, quantifying "recoverability of dropped attributes" in the model documentation as evidence of unawareness's failure.
• Verify: periodically re-check group gaps in output distributions against real outcomes — when a gap widens, first inspect which newly added features correlate with the sensitive attribute.

Related

• Same group: P4.09.2 Proxy variables re-introduce excluded attributes into models · P4.09.4 Disparate impact is judged by outcome distributions, not intent
• Adjacent: P4.02 Choosing fairness definitions · P4.03.3 Stratified evaluation
• Search terms: fairness through unawareness · redundant encodings · outcome audit