Race Discrimination in Internet Advertising: Evidence From a Field Experiment

TL;DR

The paper addresses racial bias in online advertising by conducting a large-scale field experiment on Meta's ad platform, comparing engagement for lighter versus darker complexion images and examining how salience via cropping interacts with discrimination. Using a six-test 2x2 image design and primary metric $(P_{LZ}-P_L)-(P_{DZ}-P_D)$, it finds a measurable penalty for darker skin that translates into an 11.59% higher cost to achieve the same engagement, with engagement gains more pronounced for light-skin images when salience is increased. The study also shows Meta's budget optimization tool disproportionately allocates 64% of spend to light-skin ads when enabled, amplifying observed disparities, and demonstrates that the effect persists in Brazil, suggesting robustness across geographies. The findings have implications for algorithmic fairness and regulation, highlighting how neutral systems can reinforce social biases, while also acknowledging limitations in isolating colorism from racism and in interpreting engagement as a sole proxy for broader outcomes.

Abstract

We present the results of an experiment documenting racial bias on Meta's Advertising Platform in Brazil and the United States. We find that darker skin complexions are penalized, leading to real economic consequences. For every \$1,000 an advertiser spends on ads with models with light-skin complexions, that advertiser would have to spend \$1,159 to achieve the same level of engagement using photos of darker skin complexion models. Meta's budget optimization tool reinforces these viewer biases. When pictures of models with light and dark complexions are allocated a shared budget, Meta funnels roughly 64\% of the budget towards photos featuring lighter skin complexions.

Figures (5)

• Figure 1: Illustration of the 2x2 experimental design. In the experiment, we compare two similar pictures where skin complexion differs to measure differences in how many users “Like” each ad on Instagram. We conduct a 2x2 design by then zooming in on each picture in a way that makes the skin more salient and then measuring whether this creates any disparities in “Like” rates between the two pictures, after controlling for any baseline differences in “Like” rates.
• Figure 2: Images tested. We test six sets of images in the experiment, running four ads for each of the six sets. Each set has four pictures: two similar pictures where the skin complexion differs, then two identical pictures, but zoomed in. In two of the sets, we find pictures that look similar but with different models with different skin complexions. In two of the sets, we take one picture and use Adobe Photoshop to artificially make the skin complexion look lighter. In the remaining two sets, we take one picture and use Adobe Photoshop to artificially make the skin complexion look darker.
• Figure 3: $(P_{LZ}-P_{DZ})-(P_L-P_D)$ by State. This Figure displays the measure of racial attitudes drawn from our treatment effect by viewer’s location at the state level. Darker areas correspond to areas with higher animus towards darker skin complexion advertisements.
• Figure 4: Measures of Racial Animus by State. This Figure displays various measures of racial attitudes including our treatment effect by viewer’s location at the state level. Purple cells correspond to negative correlations and green cells correspond to positive correlations. Block et al. highlight that apart from metrics that are functions of one another such as MrP and Racial Resentment, most measures of racial animus are not strongly correlated.
• Figure 5: Optimization Results: This Figure measures how advertising budget is distributed across different pictures when Meta’s budget optimization feature is turned on or off. P-value < 0.001 for Chi-Squared test comparing Amount Spent in first four bars (Optimization OFF) versus Amount Spent in last four bars (Optimization ON).