The online advertising market is strange for many reasons. It is “financialized”: real-time ad auctions let advertisers bid on standardized units of attention (impressions) based on audience data, much as a broker might bid on standardized units of oil or wheat. Except the market is essentially unregulated, the commodity is humans with agency, and the market is opaque.
See also Privacy, Algorithmic fairness, Surveillance capitalism.
Tim Hwang (2020), Subprime Attention Crisis, FSG Originals.
A short (~170 page) discussion of the online advertising market and the ways in which it resembles the subprime mortgage market in 2008. In short, the market is
This, argues Hwang, is just like the subprime mortgage crisis. Collateralized debt obligations were opaque, full of bad bonds, but also in huge demand as money needed somewhere to go. The issuing banks had no reason to stop issuing them, because the loans were sold as soon as they were issued, so they bore (apparently) no risk. And then the party stopped.
If there is a crisis of confidence in online advertising, Hwang notes, the collateral damage would wipe out numerous small websites and media outlets that rely on advertising to finance their operations.
Lambrecht, A., & Tucker, C. (2019). Algorithmic bias? An empirical study of apparent gender-based discrimination in the display of STEM career ads. Management Science, 65(7), 2966–2981. doi:10.1287/mnsc.2018.3093
Because women are a more valued demographic for advertisers, and there is more competition for their attention, prices are higher. In this case, job postings intended to be perfectly gender-neutral were seen more by men than women, because more competitive bidding for women meant that with a fixed bid, you lose the auction for women more often than for men.
Ali, M., Sapiezynski, P., Bogen, M., Korolova, A., Mislove, A., & Rieke, A. (2019). Discrimination through optimization: How Facebook’s ad delivery can lead to biased outcomes. Proceedings of the ACM on Human-Computer Interaction, 3. doi:10.1145/3359301
A much more thorough and systematic exploration of Facebook’s ad auction and targeting system. Most interestingly, they demonstrate that Facebook seems to use automated image classification to determine which audiences will respond best to the ad, regardless of the advertiser’s targeting choices; when they used images that were 98% transparent, and hence invisible to humans, there was still a massive gender disparity in delivery when the images were gendered (such as a male bodybuilder). This was not due to differing click-through rates by gender. They then demonstrate that this affects real employment ads for stereotypically gendered jobs—for example, “our five ads for janitors deliver to over 65% women and over 75% black users in aggregate”—despite all their test job ads using identical bidding and targeting strategies. This means the skew is not just because some audiences are more competitive in bidding; it must be also created by the ad platform. The conclusion includes good discussion of the policy implications (does this mean Facebook is responsible for bias in housing or employment ads, where discrimination is illegal?) and potential remedies.
Ali, M., Sapiezynski, P., Korolova, A., Mislove, A., & Rieke, A. (2019). Ad delivery algorithms: The hidden arbiters of political messaging. https://arxiv.org/abs/1912.04255
In another careful series of experiments, the same authors demonstrate political skew in ad delivery. Facebook appears to automatically analyze the landing page users are directed to by an ad, and target ads accordingly; the result is that it is easier for a Bernie Sanders ad to reach liberals than for the same ad to reach conservatives, and vice versa for a Trump ad, even if the actual users get sent to the same page. (They served partisan pages to Facebook IP addresses, but a neutral page to real users.) To reach users from the other party, you must be willing to spend much more money.
Antonio García Martínez, “How Trump Conquered Facebook—Without Russian Ads”, February 23, 2018, Wired.
Facebook (and other ad platforms) place ads based both on real-time bids by advertisers and based on their internal estimates of how likely you are to “engage” with the ad. Ads that get higher engagement are hence cheaper to run. The result:
During the run-up to the election, the Trump and Clinton campaigns bid ruthlessly for the same online real estate in front of the same swing-state voters. But because Trump used provocative content to stoke social media buzz, and he was better able to drive likes, comments, and shares than Clinton, his bids received a boost from Facebook’s click model, effectively winning him more media for less money.
Jeremy B. Merrill, “Facebook Charged Biden a Higher Price Than Trump for Campaign Ads”, October 29, 2020, The Markup.
Analyses of Facebook ads in the 2020 election suggest something similar occurred: Biden often had to pay more than Trump to place ads to similar audiences. Facebook obfuscated and denied bias, but since their bidding system is opaque, it’s impossible to determine the exact causes for the difference.