This paper argues that proxy discrimination exposes a central flaw in the engineer's defence of algorithmic screening: the assumption that predictive accuracy is sufficient to make a variable normatively admissible in high-stakes selection. The defence begins from a plausible thought: predictive systems learn from an unequal world, and engineers cannot repair the upstream institutions that produced that inequality. Yet this does not show that engineers should simply optimise for accuracy, because algorithmic systems are built through contestable choices about objectives, outcome measures, predictors, and trade-offs.
Machine-learning systems can recover protected-class information through facially neutral variables such as school attended, zip code, spending habits, or recommendation patterns facilitating proxy discrimination. Removing protected traits therefore does not prevent discrimination; it often redirects the model toward less obvious proxies. The paper develops this critique through hiring and admissions, where algorithmic screening operates within bottlenecked domains that structure access to future opportunities. In such contexts, predictive models do not merely reflect inequality but can help reproduce it by converting past disadvantage into future exclusion.
The paper argues that proxy discrimination cannot be solved by accuracy maximisation alone. Technical elimination would be the strongest response where feasible, but proxy discrimination is difficult because accuracy-seeking systems may reconstruct protected status through less visible stand-ins. Governance is therefore the ordinary practical baseline: systems must be constrained by permissible objectives and predictors, record-keeping, transparency, auditability, and contestability. In some bottlenecked domains, however, even governance may be insufficient, and structural restriction or non-deployment may be required. Accuracy may be epistemically valuable, but it cannot by itself settle admissibility where prediction allocates life-shaping opportunities.