Introduction
Most research in algorithmic fairness operates under the assumption that discrimination is a product of distributive inequality. The logic follows that if one group has less training data, lower average merit scores, or historical biases in their labels, the resulting algorithm will be biased. Consequently, the prevailing solution is to equalize these resources—merit, data, and opportunities—to make discrimination vanish.
However, this paper challenges that notion by exploring non-distributional inequality. The authors demonstrate that discrimination can emerge endogenously from the way agents interact, even when everyone starts on a perfectly level playing field. In a market where two candidates are identical in every measurable way, a market equilibrium can still force one to accept a significantly lower wage than the other.
Modeling the Hiring Market
To investigate this, the authors construct a hiring market as an extensive-form game. When a firm and a candidate match, they enter a Rubinstein-style bargaining game to split the surplus (the value created by the job). Agents utilize "threat strategies," deciding when to walk away to an "outside option," which in this case is re-entering the market to find a new match.
Crucially, the market is designed to be resource-symmetric: all agents are identical, and the game itself is symmetric. Despite this symmetry, the authors prove (Theorem 1) that there exists a range of asymmetric payoffs at equilibrium. A gap can exist where one group of identical candidates receives a significantly smaller split of the surplus than another group simply due to the "consensus" of strategies in the market.
Learning Discriminatory Equilibria
To see if these theoretical equilibria are reachable in practice, the authors model agents as learners using Follow the Regularized Leader (FTRL). They analyze whether these learners converge to Nash Equilibria (NE) and how their starting points influence the final outcome.
Through synthetic simulations, they found that agents consistently converged to Nash Equilibria within 300 steps. However, the specific equilibrium reached was highly dependent on the initial conditions of the learning algorithm—such as the first offer a firm makes or a candidate's internal valuation reference point.
This "Path Dependency" suggests that even "neutral" algorithms can "lock in" discrimination based on how they are initialized or how they break ties. If firms collectively believe a certain group will accept less, and those candidates believe they must accept less to avoid a costly walk-away, the discriminatory wage becomes a self-fulfilling prophecy.
Key Insights
- Endogenous Outside Options: In a market, your bargaining power depends on your outside options. Because your outside option depends on what other agents are doing, a feedback loop can trap a group in a low-wage state because the market consensus has devalued their alternative choices.
- Relational vs. Distributive Equality: The paper highlights that "equal merit" is not a shield against discrimination. Relational factors—power dynamics and strategy—can create gaps that standard distributive fairness metrics (like demographic parity) might detect but cannot explain or permanently fix.
Conclusion and Future Work
While the paper identifies the source of this endogenous inequality, it also points to the complexity of mitigation. The learning analysis suggests that the choice of regularizer (such as $\ell_1$, which encourages "pure" strategies) plays a role in the stability of these outcomes. Future research is needed to explore how different learning architectures or market interventions can disrupt these discriminatory consensus strategies and ensure that equal merit truly leads to equal outcomes.
Summary updated