Facebook's algorithms may lead to discriminatory ad delivery based on gender and ethnicity stereotypes — even when advertisers set parameters to be highly inclusive, a new study claims.
The study, called "Discrimination through optimization: How Facebook's ad delivery can lead to skewed outcomes," was submitted Tuesday by researchers from Northeastern University, the University of Southern California and Upturn, a Washington, D.C.-based nonprofit whose website says it "promotes equity and justice in the design, governance and use of digital technology."
The paper has not yet been submitted for peer review and said the researchers are in the process of deciding which venue would be most suited for this particular work, Upturn managing director Aaron Rieke told CNBC in an email.
The release of the study comes a week after the Trump administration charged Facebook with discrimination in its advertising practices for housing. The Department of Housing and Urban Development said in a civil complaint it was seeking damages for any person harmed by Facebook's targeted advertising policies, which until recently allowed employers and landlords to limit audiences on basis of race, ethnicity or gender.
The new study acknowledges researchers and journalists have found many ways advertisers target or exclude particular groups of users who see their ads, but says far less attention has been paid to the implications of the "ad delivery" process, which means the ways Facebook decides who should see an ad.
In Facebook's case, the study says "skewed" delivery occurs on the platform "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." The study claims an advertiser's budget and the content of the ads each "significantly" contribute to that skew. Most critically, it says there is significant skew in delivery along gender and racial lines in ads for employment and housing opportunities, even if the advertiser selects neutral categories to target Facebook users. Advertisers may not even be aware of this skewed delivery, the study says.
The researchers said in extreme cases, ads for jobs in the lumber industry reached an audience that was 72% white and 90% male, ads for cashier positions in supermarkets reached an audience that is 85% female and ads for positions at taxi companies reached a 75% black audience even though targeted audience specified by the advertiser (in this case, the researchers) was identical for all three.
"Taken together, our results paint a distressing picture of heretofore unmeasured and unaddressed skew that can occur in online advertising systems, which have significant implications for discrimination in targeted advertising," the study says.
In a statement, a Facebook spokesperson said the company has made changes to its ad platform to limit discrimination.
"We stand against discrimination in any form," the spokesperson said. "We've announced important changes to our ad targeting tools and know that this is only a first step. We've been looking at our ad delivery system and have engaged industry leaders, academics and civil rights experts on this very topic – and we're exploring more changes."
When researchers ran identical ads targeting the same audience but with varying budgets, the users who ended up seeing the ads ranged from over 55% men for ads with very low budgets to under 45% men for ads with high budgets. Skewed delivery is believed to also occur due to the content of an ad itself. Ads targeting the same audience but including a creative element referring to something that that would stereotypically be more of interest to men (like bodybuilding, researchers say) can deliver to more than 80% men, while an ad with creative elements that would stereotypically be of most interest to women (like cosmetics) could deliver to more than 90% women.
The same goes for cultural content: Ads referring to cultural content stereotypically of interest to white users, like country music, can deliver to more than 80% white users, the study says. This happens despite placing the same bid on the same audience.
The study also says researchers believe ad images are likely automatically classified by Facebook. When researchers created ads that had a high percentage of transparency so they would appear as a blank white square to humans but still included all image data, there were statistically significant differences in how ads were delivered depending on the image even though they were visually indistinguishable to the human eye.