Online ads for high-paying jobs are shown more often to men than women, according to a recent study from researchers at Carnegie Mellon University, raising questions about who is responsible for the potentially discriminatory patterns of complex algorithms.
Using test profiles of individuals interested in careers and then set to either male or female in Google Ad settings, the team of computer scientists wrote a computer program called AdFisher to visit the top 100 most popular websites for employment — to populate those profiles with an interest in careers — and then recorded the ads shown to each group.
Ads for careerchange.com, a job coaching website, advertising "$200k+ Jobs - Execs Only" were shown roughly 1,800 times to the “male” profiles and only around 300 times to the “female” profiles.
In a separate part of the study, ads for substance abuse rehabilitation and disabilities-related services were shown to profiles pre-populated with those interests by visiting websites on those topics. Google’s policy, however, prohibits remarketing, the targeting a user based on previous visits to a site, based on “health information.”
“This was a clear act of differentiation,” Anupam Datta said, one of the researchers on the study and an associate professor in Computer Science and Electrical and Computer Engineering at Carnegie Mellon. “Whether it’s the legal sense of discrimination is not for us to say, but it is certainly concerning.”
The study also found that users can sometimes successfully govern which kinds of ads they see using Google’s privacy dashboard preferences. The researchers populated two groups of profiles with an interest in dating. After removing that interest on the Google’s ad preference page for one group, the team saw a significant drop off in dating-related ads as compared to the control group.
Previous research has looked at which characteristics of online profiles correlate with different targeted ads, but this study pinpointed which characteristics actually cause those differences — in the jobs example gender is the contributing factor.
But similar to other studies, explaining which party in the chain — Google, the advertiser, the website on which the ad appeared or any number of other intermediaries — is responsible for these differences is difficult, if not impossible, Datta said.
“We don’t know who is doing the targeting, because we only have limited visibility into ecosystem,” he said. “It could be that Google’s machine learning algorithm over time may have inferred that more males were clicking on these [career services] ads and the system optimized to show them to males. If this is so, it is then unintentional automated discrimination. If the advertisers were doing the targeting, then it was possibly a human who was deciding.”
In a statement, a Google spokesperson said, “Advertisers can choose to target the audience they want to reach, and we have policies that guide the type of interest-based ads that are allowed. We provide transparency to users with Ad Settings and the ability to opt out of interest-based ads.”
Google does police bad advertising practices such as counterfeiters. In 2014 they released a report with other search and Internet companies on their efforts to remove ads for misleading weight loss programs and dietary supplements.
Google, however, doesn’t regulate how advertisers choose to market to some protected classes, such as gender. Although marketing certain products to men or women is not new in advertising, some think the pervasiveness of web advertising raises new concerns.
“The civil rights community is taking a broader view asking ‘What are we doing as a society, and how are we steering people?’” said Ali Lange, a policy analyst focused on consumer privacy at the Center for Democracy and Technology. “What opportunities are we giving to people, what types of signals are we sending on how their environment views them and how they should view themselves?’”
Although the study only looked at ads on a small number of sites, it raises questions about how to hold increasingly complicated algorithms accountable. Online advertising is just one industry from insurance rates to credit scores whose complex workings make them impenetrable to many consumers.
“What people don’t realize is, first of all, one of the implications of the complexity of the ad ecosystem is that there is absolutely no chance that anybody can do anything within the framework of notice and consent,” Frank Pasquale said, law professor at University of Maryland and author of the book “The Black Box Society: The Secret Algorithms That Control Money and Information.”
“Notice and consent” is the practice that a company discloses its practices, at least in part, and users agree to those terms.
“The second element of this is it’s the perfect smokescreen for law breaking. Frankly, if you wanted to create a legal situation that would effectively eviscerate any change of enforcement or victims finding out what happened to them, you would build an ecosystem like this,” Pasquale said.
Monitoring and interpreting such complex, opaque systems is difficult, even for Google. “In this case, the coaching service for high-paying jobs could appear slimy to some,” Michael Tschantz said, another author on the study and a researcher at the International Computer Science Institute at Berkeley. “You could make the argument that men are being negatively impacted by these ads. I think a future direction is getting inside these systems and understanding where possible cases of discrimination are happening. The future would be trying to give Google the technology to police this ecosystem more effectively.”
At what point advertising differentiation veers into legal discrimination is being hotly debated among regulators and legal scholars. The Federal Trade Commission recently held a workshop entitled “Big Data: A tool for inclusion or exclusion” and the New York University School of Law recently convened its second annual conference on Algorithms and Accountability, bringing together academics and policymakers to discuss what a regulatory framework could look like.
Pasquale points to research like this as what can inform future policy. “When studies like this come out,” Pasquale said, “we need to deeply question whether it’s worth it for the trade secrets around these difficult to understand ecosystems to be protected when the damage could be so high.”
Careerchange.com did not return a request for comment.