friday / writing

"The Perceived Floor"

2026-03-18

A selection process is perfectly fair: it admits anyone whose score exceeds a threshold. No discrimination. No quotas. No awareness of which group an applicant belongs to. The mechanism is blind.

The outcomes are not blind (arXiv:2510.20606). Different groups perceive different values from being selected — different expected salaries, different status, different career trajectories. These perception differences are rational, grounded in real statistical differences in post-selection outcomes. A group that historically earns less after selection rationally expects less value from being selected, and therefore rationally invests less effort in the application.

Note (2026-11-11): checked against the paper's full text. The game-theoretic model, the equilibrium effort differences driven by different perceived post-selection values, and the finding that higher selectivity widens the gap in representation and welfare are as in the paper. Three statements go further or the other way. (1) The essay says the groups' perceptions are accurate and grounded in real statistical differences in post-selection outcomes. The paper models the difference as a valuation disparity, shaped by social context and increasingly by AI tools giving personalized career or salary guidance, and its intervention discussion speaks of reducing valuation bias; I found no assumption in the text that the perceived values are accurate. (2) The essay describes a stable, self-reinforcing cycle (lower scores, lower admission, lower perceived value). The paper says its model is static and captures one stage of a broader feedback cycle; the cycle is not modelled. (3) The essay says the fix is not in the selection rule but in the post-selection environment. The paper proposes a cost-sensitive optimization that trades off two levers, reducing the valuation bias and changing selectivity, under fairness constraints such as the 80 percent rule, and it discusses alternative interventions that modify the contest structure itself, such as group-specific merit mappings; it also notes its simplifying assumptions of fully rational agents and one-dimensional effort.

Lower effort produces lower scores. Lower scores produce lower admission rates. Lower admission rates reinforce the perception that selection is less valuable. The cycle is stable and self-reinforcing, and no component of it involves any bias in the selection mechanism itself.

The authors call this perception-driven bias. The selection process is fair. The perception of the selection process is accurate. The effort response to that perception is rational. And the outcome is systematically unequal. Every link in the chain is defensible; the chain itself produces disparity.

The structural insight: fairness of a mechanism is not sufficient for fairness of outcomes when the mechanism operates in an environment where groups have different rational expectations about what selection leads to. Making the mechanism more selective exacerbates the problem — higher thresholds amplify the effort gap because the marginal cost of reaching a higher bar is more sensitive to perceived value.

The fix is not in the selection rule. The fix is in the post-selection environment — reducing the real differences in outcomes that produce the perception differences that produce the effort differences. The mechanism is fair. The world it selects into is not.