Strategic Costs of Perceived Bias in Fair Selection

1. Problem Statement

The paper addresses a paradox in meritocratic systems (e.g., university admissions, hiring): despite formal neutrality and identical selection criteria, significant disparities persist across race, gender, and class. While traditional explanations focus on structural barriers (unequal access to resources) or statistical discrimination (biased beliefs by selectors), this paper investigates a perception-driven bias.

The core hypothesis is that candidates from different socioeconomic groups may possess different perceived values of being selected ($v$). These perceptions are shaped by social context, historical inequalities, and increasingly, algorithmic tools (like LLMs) that provide biased career or salary guidance. Even if the selection process is perfectly fair, rational agents will strategically adjust their effort levels based on these perceived returns. If Group A perceives a lower return on success than Group B, Group A will rationally exert less effort, leading to lower representation and welfare, thereby perpetuating inequality through a "fair" mechanism.

2. Methodology

The authors develop a game-theoretic model combining contest theory with algorithmic fairness concepts.

• The Model:

  • Agents: A population of $n$ agents divided into two groups: an advantaged group $G_1$ (size $(1-\alpha)n$) and a disadvantaged group $G_2$ (size $\alpha n$).
  • Valuations: Each agent $i$ has a private valuation $v_i$ for being selected. $v_i \sim p_\ell$ for group $\ell$. The model assumes systemic disadvantage via a scaling factor $\rho \in (0, 1]$, where $p_2(v) = \frac{1}{\rho}p_1(\frac{v}{\rho})$. This implies $E[v_2] = \rho E[v_1]$.
  • Effort and Merit: Agents choose effort $e_i$ to maximize expected payoff. The observable score is $s_i = e_i + a_i$ (where $a_i$ is initial ability). Selection is based on the top $k = cn$ scores.
  • Payoff: $\pi_i = \mathbb{I}(\text{selected}) \cdot v_i - e_i$.
  • Goal: Characterize the Nash Equilibrium (NE) strategies and analyze metrics like representation ratio ($r_R$) and social welfare ratio ($r_S$).

• Analytical Approach:

  • The authors analyze the large-agent limit ($n \to \infty$) to derive tractable closed-form solutions, as finite-$n$ NE policies are complex and often asymmetric.
  • They prove the existence and uniqueness of a symmetric NE in the limit and establish the convergence rate of finite-$n$ policies to this limit ($O(\sqrt{\log n / n})$).
  • They derive explicit formulas for key metrics under uniform distributions and propose a cost-sensitive optimization framework to design interventions.

3. Key Contributions

1. Modeling Perception-Driven Bias:
   The paper introduces a novel framework where bias is not in the selection rule but in the valuation distribution of the candidates. It demonstrates how differences in perceived post-selection value ($\rho$) alone can drive rational differences in effort, creating disparities even in a "blind" meritocratic contest.

2. Equilibrium Characterization (Theorem 4.1):
   The authors prove that in the large-$n$ limit, the NE policy for both groups converges to a single threshold strategy $s(v, a)$.

   • Agents exert effort only if their combined valuation and ability ($v+a$) exceed a global threshold $t$.
   • The threshold $t$ is the unique solution to $(1-\alpha)F_1(t) + \alpha F_2(t) = 1-c$, where $F_\ell$ are the CDFs of the groups.
   • This implies that in equilibrium, the strategic environment becomes effectively uniform across groups, despite the underlying valuation asymmetry.

3. Micro-to-Macro Metric Analysis:
   The paper derives closed-form expressions for:

   • Representation Ratio ($r_R$): The ratio of selection rates between groups.
   • Social Welfare Ratio ($r_S$): The ratio of average net payoffs.
   • Institutional Revenue ($RV$): The average merit of selected agents.
   • Key Finding: These metrics exhibit non-linear behavior and sharp thresholds. For instance, if the bias $\rho$ drops below a critical value relative to selectivity $c$, the disadvantaged group may completely stop participating ($r_R = 0$).

4. Intervention Design Framework:
   The authors formulate a constrained optimization problem to minimize the cost of interventions required to meet fairness targets (e.g., the "80% rule" where $r_R \ge 0.8$).

   • Levers: Institutions can either increase selectivity ($c$) or reduce valuation bias ($\rho$).
   • Trade-offs: The analysis reveals that for high disparity targets, increasing access (raising $c$) is more cost-effective, whereas for narrowing smaller gaps, improving perceived value (raising $\rho$) is better.

4. Key Results and Findings

• The "Perception-Driven" Mechanism: Disparities are not just a result of structural barriers but are rational responses to perceived low returns. If a group believes the reward for success is low (low $\rho$), they rationally under-invest in effort, leading to lower selection rates.
• Sharp Thresholds: There is a critical threshold for $\rho$ (specifically $\rho_c = 1 - \frac{c}{1-\alpha}$).
  • If $\rho < \rho_c$, the disadvantaged group exerts zero effort, resulting in zero representation.
  • If $\rho > \rho_c$, representation and welfare increase non-linearly with $\rho$.
• Impact of Selectivity: Increasing selectivity (lowering $c$) exacerbates disparities. In highly competitive environments (low $c$), even small differences in perceived value lead to massive gaps in representation.
• Intervention Insights (JEE Case Study): Using data from the JEE Advanced 2024 exam, the authors calibrated the model. They found that to achieve an 80% representation ratio:
  • Simply increasing the selection rate ($c$) is often cheaper than trying to fix the perceived value bias ($\rho$) when the gap is large.
  • However, as the gap narrows, fixing the perception bias becomes the more efficient path.
• Revenue vs. Fairness: The paper shows a trade-off: increasing selectivity ($c$) to improve fairness (by lowering the threshold for the disadvantaged group) actually decreases the average merit (revenue) of the selected pool because it lowers the global threshold $t$.

5. Significance and Implications

• Bridging Rational Choice and Structural Inequality: The paper bridges two schools of thought. It shows that structural inequalities (perceived value) can manifest as rational individual choices (effort), creating a feedback loop that sustains inequality even without explicit discrimination.
• Role of AI: It highlights a critical risk of AI tools: if LLMs provide biased salary or career advice, they effectively alter the parameter $\rho$ in the model, potentially causing rational agents to disengage from meritocratic processes, thereby validating the bias they were warned about.
• Policy Design: The work provides a quantitative toolkit for institutions. It suggests that "fair" selection processes are insufficient; institutions must actively manage the perceived value of selection for different groups (e.g., through mentorship, transparent salary data, or targeted support) to prevent the collapse of participation from disadvantaged groups.
• Theoretical Advancement: It extends contest theory to multi-group settings with asymmetric valuations, providing rigorous proofs for equilibrium existence and convergence in large populations, which had previously been an open challenge.

In summary, the paper argues that fairness is not just about the selection rule, but about the incentives created by the perceived value of the outcome. Without addressing the perception gap, meritocratic systems will naturally reproduce existing disparities through the rational behavior of candidates.