Fighting discrimination with reputation: The case of online platforms

Reputation Systems Mitigate Ethnic Discrimination on Ridesharing Platforms

On many major online marketplaces, the discretion of participants often leads to discrimination based on race, ethnicity, or gender. While these platforms are built around reputation systems—aggregating reviews and ratings into signals of individual performance—it is unclear if these signals actually correct for social biases or merely mirror them. A new study of a large French ridesharing platform suggests that reputation can indeed act as a corrective tool.

The researchers find that while new minority drivers initially earn significantly less than their nonminority counterparts, this gap nearly vanishes as they accumulate reviews. The study reveals that this isn't just a passive observation of experience. Rather, minority drivers actively invest in their reputations. They do this by cutting introductory prices and exerting extra effort to overturn skeptical passenger beliefs.

The revenue gap at entry

The study identifies a significant economic disadvantage for minority drivers entering the BlaBlaCar marketplace. The authors report that ethnic minority drivers with little or no reputation earn 11.6% less revenue than otherwise similar nonminority drivers. This disparity is not driven by price alone. Although minority drivers post slightly lower fares, their listings attract fewer views and sell fewer seats.

This disadvantage is heavily concentrated in the early stages of a driver's career. As shown in, the gap in both sold seats and revenue per kilometer is widest for drivers in the first two deciles of review counts.

However, the authors find that the disparity is not a permanent fixture of a driver's identity. As drivers accumulate more reviews, the ethnic differences in outcomes become statistically insignificant. This suggests that the reputation system provides a pathway to parity.

A model of strategic reputation building

To understand why this convergence happens, the authors develop a structural model of "career concerns." This model treats the driver's career as a dynamic optimization problem. Current actions—specifically pricing and effort—are chosen to maximize lifetime payoffs by influencing future passenger beliefs.

The mechanism operates through three primary layers:

1. Belief Formation: Passengers do not start with a blank slate. They hold "priors"—initial beliefs about a driver's quality—that are influenced by group identity. The authors find that these priors are often inaccurately pessimistic for minority groups.
2. Strategic Pricing: Because reviews are valuable for correcting these pessimistic priors, drivers have an incentive to "buy" reputation. The authors report that minority entrants set introductory prices 7.2% below their long-term profit-maximizing level. In comparison, nonminority entrants only discount by 4.7%.
3. Strategic Effort: Beyond price, drivers use service quality to boost ratings. The authors find that minority entrants exert more effort than nonminority entrants. They do this to ensure their early reviews are high enough to trigger a belief revision.

This process is visually supported by .

It shows that both pricing discounts and the share of 5-star ratings peak during the first 15 reviews. These behaviors decline as the marginal benefit of an additional review diminishes.

Causal evidence from a railway strike

A central challenge in studying discrimination is determining whether observed gaps are caused by actual differences in driver quality or by passenger bias. To isolate the causal effect of reputation, the authors exploit a natural experiment: the 2018 French railway strike.

Because BlaBlaCar and trains are close substitutes, the strike created a massive, exogenous demand shock. Drivers operating on strike days saw booking requests increase sixfold. This led to a rapid accumulation of reviews. The authors used a difference-in-differences design to compare drivers who drove during the strike to those who did not.

The results, visualized in, show a sharp spike in both seats sold and revenue on strike days.

Crucially, the authors report that the post-strike revenue gain was 61% larger for minority drivers than for nonminority drivers. This proves that reviews are most valuable for the very drivers passengers initially judge with the most skepticism. The strike essentially acted as an "accelerated learning" event. It allowed minority drivers to bypass the slow grind of reputation building.

The limits of information

While the findings are optimistic about the power of reputation, the authors highlight several critical caveats. First, the reputation system is not a panacea. The study notes that the cost of overcoming prejudice is borne entirely by the minority drivers. They must sacrifice immediate revenue through discounts and incur higher costs through extra effort.

Second, the effectiveness of the system is highly dependent on the design of the rating scale. The authors ran counterfactual simulations to test how different "signal precisions" affect discrimination. They found that while a "sharper" (more informative) rating system helps the market correct biased beliefs faster, it cannot eliminate the gap entirely.

Even under ideal conditions, a residual "statistical-discrimination floor" remains. This refers to a persistent gap caused by passengers rationally applying group-level averages to individuals when information is still imperfect. This floor represents the unavoidable uncertainty inherent in any marketplace.

Finally, the study does not explore the root causes of the biased priors themselves. The model can quantify the cost of being underestimated, but it cannot explain why passengers hold those pessimistic beliefs in the first place.

The verdict on platform design

The verdict is that reputation systems are powerful, but they are reactive rather than proactive. They allow marginalized participants to "prove" their way out of a disadvantage. However, they do not prevent the disadvantage from occurring at the moment of entry.

For platform engineers and designers, the takeaway is clear. The speed of bias correction is a tunable parameter. The authors demonstrate that increasing the informativeness of the rating system can narrow the discrimination wedge by roughly 20%. If you are building a marketplace where identity is visible, your goal should be to maximize the ability of reviews to overwrite biased priors.

Figures from the paper