Recently, a research team at Carnegie Mellon led by Prof. Annupam Datta set out to test their hypothesis that high-paying executive jobs advertised via Google Adwords would be served to men more often than to women. (Washington Post wrote all about it here.) They set up an experiment in which they created a series of fake user accounts with the exact same data, save that half of them were male and half were female. They visited the same series of job search sites with each account and then recorded the ads that they were served. What they found was shocking: of the 2,170 ads for high-paying executive jobs they recorded, 85% were shown to the “male” accounts – meaning the ads were served to “male” profiles at nearly 6 times the rate they were served to “female” profiles.
While there is rampant speculation around the various reasons for this result, the experiment essentially proves that blame rests squarely on the advertisers themselves, many of whom had to have specifically targeted men to receive these ads for this result to be possible. Amid cries of “but the algorithm is so complex, we’ll never know!”, I thought it would be helpful to break down the different ways that Google allows its advertisers to target their ads to examine the evidence.
Contextual Targeting: Google shows ads on websites that have content that aligns with keywords the advertisers have provided. In this experiment, all of the profiles started with clean search history and visited the same websites, so this type of targeting would not have resulted in users seeing different ads.
Placement Targeting: Placement allows advertisers to hand-select the websites where they want their ads to be displayed. Again, since all profiles visited the same site, this type of targeting wouldn’t change what type of ads each gender saw.
Remarketing: Some websites have tags embedded in their code that allow ads to be served to anyone who has visited that particular website. Since all profiles visited the same site, this would not affect the outcome of the experiment.
Interest Categories: A form of behavioral targeting, this allows advertisers to choose to serve ads to individuals who have expressed specific interests, like “sports”. Since all profiles in this experiment were identical save gender identification, this type of targeting would not affect the result.
Topic Targeting: Ads are served on websites that fit into the categories of the users’ expressed interest (like ESPN for “sports”). Since all profiles visited the same sites, this would not come into play.
Geographic and Language Targeting: Advertisers can serve ads to individuals in a specific location and/or who have selected specific language preferences. Since all profiles were identical and the study was conducted in the same location, this type of targeting would not influence the result.
Demographic Targeting: Google allows advertisers to serve ads to limited age and gender ranges. On the gender side, the options provided are “male”, “female”, and “unknown”. Advertisers can select one, two, or all three options. Given the parameters of the experiment, this is the only type of targeting that could have affected the outcome. In order for 85% of the ads to have been served to the “male” profiles, a large majority of the advertisers would have had to have selected only “male” or only “male” and “unknown” to target their ads to. The only other plausible explanation is that one advertiser targeted this way, and ran such a significant majority of the ads that it heavily skewed the results of the experiment.
Google, as a company, is incentivized to make money – which means it targets ads according to advertiser specifications, and then optimizes that service based on what types of people are engaging with the ads. When ads are mostly served to males, then mostly males will engage with them, which means they will be served even more heavily to males. This algorithm ensures that advertisers get the most efficient spend for their money, but also can’t control for the advertiser’s intentions when setting their original targeting inputs. So at least in this case, when rampant sexism in the digital targeting of executive job ads is exposed, the guilt likely rests with the companies and agencies placing the ad buys.