Source-linked AI summary
Discrimination through optimization: How Facebook's ad delivery can lead to skewed outcomes
Muhammad Ali, Piotr Sapiezynski, Miranda Bogen, Aleksandra Korolova, Alan Mislove, Aaron Rieke
TL;DR
Prior work had examined advertiser targeting more than platform ad delivery, leaving it unclear whether delivery systems could skew audiences without advertisers’ intent. The paper experimentally studies Facebook’s delivery process and finds that budgets, ad content, and automated relevance estimates contribute to demographic skew, including in employment and housing ads with identical targeting. The authors conclude that delivery optimization itself must be considered when evaluating potentially discriminatory digital advertising.
Problem
Research had found discriminatory targeting or exclusion, but comparatively little evidence addressed whether platform ad delivery independently skews audiences based on demographics.
Method
The authors run Facebook campaigns and controlled ad experiments varying budgets, creative components, and image transparency to distinguish market effects from delivery optimization.
Results
Facebook delivery skewed audiences through budgets, ad creatives, and likely automated relevance classification, including substantial gender and racial skew in employment and housing ads with identical targeting.
Takeaways & Limitations
Ad-platform delivery optimization can independently shape potentially discriminatory outcomes, so evaluation of digital advertising must consider delivery as well as advertiser targeting.
Takeaways & Limitations
The employment and housing findings concern the particular ads studied and cannot establish how such ads are delivered in general.
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
Online advertising research has focused more on advertiser targeting than on how platforms deliver ads. This paper investigates Facebook’s delivery optimization and finds that budgets, ad creatives, and automated relevance estimates can produce demographic skew, including in employment and housing advertising.
- Facebook ad delivery selects users using factors including advertiser budgets, ad performance, and predicted ad relevance.
- The study addresses whether delivery alone can skew which users see ads, independently of advertisers’ targeting choices.The authors ran dozens of campaigns and hundreds of ads, generating millions of impressions while spending over $8,500.
- Over 55% men versus under 45% men received identical ads targeting the same audience when budgets varied from very low to high.This demonstrates skew arising from market effects alone.
- Ad creatives alone produced heavily skewed delivery: bodybuilding ads reached over 80% men, cosmetics ads over 90% women, hip-hop ads over 85% Black users, and country-music ads over 80% white users.These ads used the same target audience and bid.
- The ad image significantly affected delivery, including when headlines and text stereotypically appealed to men but the image stereotypically appealed to women.
- Visually indistinguishable images with over 98% transparency received statistically different delivery, indicating automated image classification and relevance estimates can skew delivery from the start.The result suggests skew can arise from Facebook’s automated relevance estimate rather than viewers’ interactions.
- Real employment and housing ads also showed substantial demographic skew despite identical targeting, with extreme cases reaching 90% male, 85% female, or 75% Black audiences.The reported examples included lumber jobs, supermarket cashier positions, taxi-company positions, and housing ads reaching over 72% or over 51% Black users.
- The findings identify platform delivery optimization and market effects as mechanisms that can create potentially discriminatory outcomes beyond advertisers’ targeting choices.The paper highlights implications for housing and employment advertising and calls for careful consideration by regulators, lawmakers, and platforms.
2 BACKGROUND
Online advertising platforms separate ad creation from ad delivery, using auctions and optimization to decide which eligible users see ads. Facebook combines advertiser inputs with bids, performance estimates, and relevance judgments, creating mechanisms through which delivery can skew beyond targeting choices.
- Ad creation: Ad creation lets advertisers specify creative content, audience targeting, and bidding strategy before submitting an ad for platform review.Creative content includes headlines, text, and images or videos; advertisers also provide a destination link and bid or bid cap.
- Ad delivery: Ad delivery runs an auction among ads whose audiences include the current user, selecting which advertisement appears in an available slot.Platforms may also avoid repeating the same advertiser’s ads to a user in quick succession.
- Audience selection: Advertisers can target users through demographics, personal information, tracking pixels, and lookalike audiences, and can combine these options.Facebook’s targeting system includes more than 1,000 well-defined attributes and hundreds of thousands of free-form attributes.
- Ad delivery: Facebook’s winning-ad value combines bid, estimated action rates, and ad quality and relevance rather than relying on bids alone.Estimated action rates reflect general engagement, while ad quality and relevance estimate how interesting or useful an ad may be to a particular user.
- Paper focus: The paper examines single-site platforms and separates market effects from optimization effects as mechanisms through which delivery may become skewed.The authors focus on Facebook and note that applicability to broader web advertising remains for future investigation.
3 METHODOLOGY
The methodology varies selected ad features while controlling target audiences, timing, and auction competition, then measures demographic composition in delivered audiences. Gender is measured through Facebook delivery breakdowns, while racial delivery is inferred from DMA-based custom-audience construction.
- Experimental design: The experiments vary one ad feature at a time and measure how that change skews the users reached within a controlled target audience.Random, mutually exclusive PII-based custom audiences help control which auctions ads enter and prevent experimental instances from competing with one another.
- Delivery measurement: Facebook’s Marketing API is queried every two minutes for delivery statistics broken down by demographic or geographic attributes.The analysis uses reach, defined as unique users shown the ad, and calculates the fraction of men from male and female reach.
- Racial measurement: Racial delivery is inferred by using Designated Market Areas as a proxy because Facebook’s API does not provide racial breakdowns.North Carolina voter records are partitioned into matched DMA sets and uploaded as custom audiences, enabling delivery by DMA to be mapped back to race.
- Campaign settings: Experiments generally target U.S. adults across genders, use traffic optimization with a single relevant image and text, and run most ads on a $20-per-day budget.Ads are run simultaneously within experiments to control for time-of-day effects, and delivery is collected through the Facebook Ad API.
- Measurement scope: The analysis restricts gender comparisons to self-reported female and male users and racial comparisons to self-reported Black and White users.The authors acknowledge that skew involving non-binary users and other races is not reported.
4 EXPERIMENTS
Experiments show that Facebook’s ad delivery can skew audiences even when targeting and bidding are held constant, with budget, creative content, images, and automated classification each contributing. The skew extends to real employment and housing ads across gender and racial lines.
- 4.2 Ad creative effects on ad delivery: With identical bidding and gender-agnostic targeting, bodybuilding ads reached over 75% men on average, whereas cosmetics ads reached over 90% women.The observed audience differences occurred despite no advertiser-specified difference in budget or target audience.
- 4.2 Ad creative effects on ad delivery: The ad image alone could dramatically change delivery: adding an image produced the major shift, while restoring text or headlines alongside it did not significantly alter the skew.Base bodybuilding and cosmetics ads reached 48% and 40% men, respectively, before the image was introduced.
- 4.2 Ad creative effects on ad delivery: Swapping images reversed gender proportions within an hour, without a significant change in gender-specific click-through rates, suggesting delivery changed independently of observable user response.The image swap occurred after six hours and 32 minutes of campaigning.
- 4.3 Source of ad delivery skew: Nearly transparent masculine and feminine images received significantly different delivery, indicating that Facebook likely classified images automatically from the beginning of campaigns.The images retained their data but appeared blank or indistinguishable to human viewers; the authors do not know exactly how classification works.
- 4.4 Impact on real ads: Real employment and housing ads with identical targeting and bidding reached sharply different demographic audiences, including employment ads exceeding 90% men and 70% white users and housing ads ranging from over 72% to as little as 51% Black users.The employment experiments varied ad creative and destination link while holding targeting, bidding, timing, and competition conditions constant; the housing results showed racial but little gender skew.
5 CONCLUDING DISCUSSION
The paper shows that Facebook’s ad delivery optimization can independently steer broadly targeted ads toward skewed demographic subsets, extending concerns beyond advertiser-controlled targeting. The authors connect this finding to legal, transparency, and policy implications while cautioning that experimental results do not support broad conclusions about all ads.
- Findings: Facebook’s delivery process can significantly alter an ad’s audience based on ad content, even when budget and target audience remain constant.The authors used public voter-record data to show differential delivery to specific audience segments.
- Findings: Lower daily budgets were associated with delivery to fewer women, replicating prior evidence that market and pricing dynamics can produce differential outcomes.
- Findings: Skewed delivery can begin at the start of an ad’s run and is likely automated rather than merely a reflection of early user feedback.Transparent images that appeared identical to humans but differed to automatic classifiers produced skewed delivery.
- Implications: Housing and employment ads showed skewed delivery across legally protected categories, making such outcomes potentially discriminatory even with broad and inclusive targeting.The paper distinguishes these cases from cosmetics or bodybuilding ads, whose gender skew is unlikely to have legal implications.
- Limitations: The study cannot support broad conclusions about how Facebook’s delivery process affects ads generally.The authors note that observed patterns, such as predominantly white and male delivery for lumberjack ads, may not hold for all ads in that category.
- Policy implications: Because delivery can independently generate discrimination, restricting advertiser targeting alone will not address delivery-based discrimination.The authors argue that broader target audiences could increase the practical impact of delivery mechanisms.
- Policy implications: The authors call for greater transparency about delivery algorithms and statistics, especially for housing, credit, and employment advertising.They state that Facebook’s existing transparency efforts do not yet permit real-world analysis of delivery impacts.
ERRATA
The erratum corrects a reported delivery figure and clarifies that white-user audience fractions are estimates.
- The corrected figure for ads delivered outside predefined DMAs is approximately 10%, replacing the previously reported approximately 40–50%.
- The figure axis labels were changed to “Estimated fraction of white users in the audience.”
- The confidence-interval calculation method was changed from normal approximation to the method described in the revised version.
APPENDIX
The appendix reports a multiple-hypotheses check for the employment-ad experiment.
- The appendix aggregates five image variants for each of 11 job postings and compares male and estimated white-user fractions between job pairs.This produces 55 tests for assessing whether apparent differences could reflect multiple-hypothesis testing.
- The analysis uses aggregated job-level comparisons rather than treating the five image variants for each posting as separate tests.