A selection process is designed to be fair: candidates are evaluated purely on merit, blind to group membership. No bias in evaluation. No structural advantage. The mechanism is clean.
Note (2026-11-10): I checked the abstract (current and first version) and searched the full text of arXiv 2510.20606 (Celis, Huang, Sohoni, Vishnoi). The model and the perception-driven bias match the essay. The essay ends with fairness being necessary but not sufficient, and does not mention that the paper goes on to propose and solve an intervention: a constrained optimization in which an institution trades off raising selectivity against reducing the valuation bias under fairness constraints such as the 80 percent rule, in closed form for linear costs, which the abstract says can reduce disparities without compromising institutional goals. The essay also does not mention the abstract's note that the perceived value can be shaped by AI-powered career or salary guidance, or that the equilibrium characterization is for the large-agent limit. The essay's line that the cleaner the meritocracy the more precisely it transmits the signal is its own.
The disparity appears anyway, because the mechanism is not the only thing that matters. When different groups perceive different post-selection values — different expected returns from being selected, different likelihoods of success in the post-selection environment — they rationally adjust their effort in response (arXiv:2510.20606). Group A, perceiving high returns, invests heavily. Group B, perceiving lower returns, invests less. The merit-based evaluation then correctly identifies Group A as higher-performing. The outcome gap is real, the evaluation is unbiased, and the disparity is driven entirely by rational responses to perceived future conditions.
The fairness mechanism becomes a channel for inequality transmission. It does not create the disparity — the disparity originates in differential post-selection perceptions. But the mechanism faithfully propagates it: by evaluating effort-dependent merit, it converts perception differences into outcome differences. The cleaner the meritocracy, the more precisely it transmits the upstream signal.
This is structurally different from discrimination. There is no biased evaluator to identify and remove. The mechanism works exactly as designed. The problem is that “works as designed” includes amplifying any pre-existing differences in how groups perceive their prospects. A fair filter applied to an unfair world produces unfair results, and the fairness of the filter is precisely what makes it an effective transmitter of the world's unfairness.
The implication: fairness in the mechanism is necessary but not sufficient. The mechanism's fairness guarantees only that the output accurately reflects the input. If the input is shaped by inequality, the output will be too — with the mechanism's stamp of legitimacy.