[Paper Review] Discrimination through optimization: How Facebook's ad delivery can lead to skewed outcomes

Muhammad Ali, Piotr Sapieżyński+4 more|arXiv (Cornell University)|Apr 3, 2019

Consumer Market Behavior and Pricing32 references74 citations

TL;DR

The paper demonstrates that Facebook’s ad delivery can skew who sees ads for employment and housing based on ad budget and creative content, even with inclusive targeting, revealing mechanisms by which platform optimization and market effects drive discriminatory outcomes.

ABSTRACT

The enormous financial success of online advertising platforms is partially due to the precise targeting features they offer. Although researchers and journalists have found many ways that advertisers can target---or exclude---particular groups of users seeing their ads, comparatively little attention has been paid to the implications of the platform's ad delivery process, comprised of the platform's choices about which users see which ads. It has been hypothesized that this process can "skew" ad delivery in ways that the advertisers do not intend, making some users less likely than others to see particular ads based on their demographic characteristics. In this paper, we demonstrate that such skewed delivery occurs on Facebook, due to market and financial optimization effects as well as the platform's own predictions about the "relevance" of ads to different groups of users. We find that both the advertiser's budget and the content of the ad each significantly contribute to the skew of Facebook's ad delivery. Critically, we observe significant skew in delivery along gender and racial lines for "real" ads for employment and housing opportunities despite neutral targeting parameters. Our results demonstrate previously unknown mechanisms that can lead to potentially discriminatory ad delivery, even when advertisers set their targeting parameters to be highly inclusive. This underscores the need for policymakers and platforms to carefully consider the role of the ad delivery optimization run by ad platforms themselves---and not just the targeting choices of advertisers---in preventing discrimination in digital advertising.

Motivation & Objective

• Assess whether Facebook’s ad delivery alone can produce skewed delivery across demographic groups regardless of targeting parameters.
• Quantify the impact of budget, ad creative, and images on delivery skew along gender and race.
• Investigate whether automated relevance classification contributes to initial delivery skew.
• Demonstrate real-world implications for employment and housing ads and regulatory considerations.

Proposed method

• Run dozens of Facebook ad campaigns with controlled targeting and varying budgets and creatives.
• Use the Facebook Marketing API to collect delivery statistics at two-minute intervals across demographic dimensions.
• Develop methods to infer racial delivery using DMA-level location proxies and publicly available voter race data.
• Compare delivery skew across budget levels, ad creative content, and image components while holding targeting constant.
• Conduct experiments with real employment and housing ads to observe skew in protected-class representation.

Experimental results

Research questions

• RQ1 Does ad delivery skew occur under market/optimization effects even when targeting is neutral or inclusive?
• RQ2 How do budget, creative content, and imagery affect the demographic composition of the delivered audience?
• RQ3 Can automated relevance assessments contribute to initial delivery skew independent of user interaction?
• RQ4 Do real-world employment and housing ads exhibit significant demographic skew in delivery under identical targeting?

Key findings

• Skewed delivery occurs due to market effects alone, with audience composition varying with budget (e.g., >55% men at very low budgets, <45% men at high budgets).
• Ad creative content strongly influences delivery skew, with certain stereotypes driving >80% male or >90% female delivery under the same targeting.
• Ad images independently affect delivery; swapping headlines/text/images alters delivery patterns, indicating image-driven relevance signals.
• Automated image classification by Facebook is likely contributing to skew from the start of ad runs, as visually indistinguishable images yield different delivery.
• Real employment and housing ads show substantial skew under same targeting, e.g., jobs reaching 72% white and 90% male in some cases, or 85% female audience for cashier ads; housing ads skew with >72% Black in some variants and >51% Black in others.
• The findings suggest ad-delivery optimization plus market effects can create discriminatory outcomes beyond targeting controls, prompting regulatory and platform considerations.