A model's outcomes differ across demographic groups. What can architecture actually do about it?

The first difficulty: fairness is not one thing

Several reasonable definitions exist — equal outcome rates, equal error rates, equal treatment of similar individuals — and they are mathematically incompatible except in degenerate cases. Satisfying one means failing another.

So the first requirement is a stated definition chosen for the context, with the reasoning recorded. A system optimising an unstated notion of fairness cannot be assessed, and the choice is a policy decision rather than a technical one.

What can be done at each stage

• Data: examine representation and label quality across groups. Historical labels encode historical decisions, so a model trained on them reproduces past disparity faithfully — which is the most common origin and the least fixable by modelling technique.
• Training: constraints and reweighting, which trade aggregate performance for distributional properties.
• Post-processing: group-specific thresholds, effective and legally fraught in many jurisdictions.
• Deployment: human review for affected decisions, an appeal route, and explanation.

What architecture specifically contributes

• Measurement infrastructure. Disparity cannot be managed without segmented performance monitoring in production, which requires the attributes to measure against — and collecting them raises its own privacy question that must be resolved deliberately rather than avoided.
• Contestability: a path for an affected person to seek review, which is frequently the most meaningful control and the one most often absent.
• A fallback to human decision where the model's confidence is low or the case is unusual.
• Monitoring for drift in disparity, since a system fair at launch can become unfair as the population changes.

The honest position

Removing the protected attribute does not remove the bias. Correlated features reconstruct it, and removing the attribute also removes the ability to measure disparity — which makes the system look neutral while behaving otherwise. Measure it, decide the definition, state the trade-off, and monitor it — that is what architecture can support; the rest is a policy decision that must be made explicitly and by someone accountable.