Ensuring equity in online advertising for employment, housing, and credit
by Jamie Morgenstern, Assistant Professor in the Paul G. Allen School of Computer Science and Engineering at the University of Washington.
Online advertising systems have received a great deal of recent scrutiny regarding their systems’ displaying of high-stakes advertisements to different demographic groups at wildly different rates. Examples include advertising for high-wage jobs being served to men at higher rates than women [1], arrest record search sites being shown at higher rates when searching for black-identifying names than white-identifying names [2], showing hiring advertisements less often to older users [3], and showing housing advertisements in a way which strongly correlated with protected categories such as sex, family status, disability, national origin, and other protected characteristics [4,5,6]. Some of these systems explicitly allow advertisers to name demographics they want to see (or not see) their ads and some allow the advertisers to select their audience based on peoples’ behavior on the platform or elsewhere online.
Platforms that do not allow advertisers to explicitly target demographic groups can still allow intentional or unintentional redlining: a user’s behavior or information on a platform can be strongly correlated with their demographics. Thus, removing explicit demographic targeting from online advertising will not necessarily ensure users from different demographics will see an ad with equal frequency.
Even when platforms allow no targeting for advertisements in a given
domain (say, for example, housing advertisements), the ability of other advertisers in possibly lower-stakes industries to target can still lead to housing advertisements being shown at wildly different rates to different demographics.
To give a very simple example of how this might occur, suppose that an
advertiser selling running shoes targets people between the ages of 25 and 40 who the platform identified as women. This advertiser might be willing to pay $.10 for each time the platform displays this shoe ad to someone in this population, and $.02 each time the platform displays it to anyone else. Suppose a landlord with a new apartment building might want to advertise in
the same timeframe as the shoe company. They might be willing to pay $.05 for each person who sees their ad. Assume the advertising platform shows each user a single ad, and chooses to show the one which would pay more for their impression. This will result in targeted people seeing ads from the shoe company, and un-targeted people seeing ads for the new apartment building. So, the willingness of a retail company to pay more to show their advertisement to certain demographics results in that demographic seeing fewer housing advertisements.
Many of the worrisome examples outlined here belong to industries with
higher than normal legal scrutiny regarding how the industries’ practices reach people from different demographics. The Civil Rights Act provides extra tools to address unequal access to organizations involved in voting, education, public accommodations, and employment. The Equal Credit Opportunity Act provides similar levers in the context of consumer lending. Unfortunately, legal avenues are usually very slow in their ability to affect change, and technologies that underly personalization systems are rapidly evolving.
Therefore, relying on our legal system to “sort out” which forms of personalized advertisements are legal and which are not is unlikely to reduce the differing rates at which impactful advertisements are shown to different demographics.
What, then, can we do about targeted advertising systems to ensure
ads which are “high-stakes” are shown to every demographic equally?
As outlined above, it is not sufficient to ensure that high-stakes advertisements are not given the ability to explicitly target demographic groups, as these groups can be uncovered with targeting through other characteristics. It is also not sufficient to remove all targeting from high-stakes advertisers, as in most existing systems, these advertisers will still need to compete with advertisers who can target. Recent research has proposed bidding strategies that advertisers might employ to ensure their ads are displayed equitably [7]. However, such approaches leave the responsibility of displaying advertising equitably to advertisers, who are likely far from expert in either the technical inner workings of online ad systems or the legal consequences for their ads reaching only certain demographic groups.
In my opinion, the only remaining possible intervention which works within existing targeted advertising frameworks is to:
1. Announce the percentage of advertisements that will be allocated to housing, lending, and employment that each person using a system will be served, and then
2. Categorize every advertisement as either pertaining to housing, lending, employment, versus those categories where targeting is allowed (advertisements for sneakers, for example, might strike most people as an acceptable place for targeting leading to some demographics seeing such ads at higher rates than others). Finally,
3. Run separate ad markets for each of these categories, and do zero targeting within the categories of housing, lending, and employment.
This shift, from every advertising opportunity being available to every advertiser willing to pay enough, will ensure, no matter how advertisers behave once categorized, that these high-stakes ads will be shown to each demographic equally. It will still allow for companies who see large revenue streams stemming from targeted advertisement to capitalize on that revenue, in the domains where such targeting is considered acceptable.
Several recent projects have designed targeting advertisement platforms that do not introduce unfairness when advertisers place fair bids [8, 9], encouraging the users of a platform to express their preferences over categories of advertisements [8] rather than having the percentage of advertisements prespecified by the platform as I suggest above. These approaches have tools and perspectives that are definitely worthy of consideration and implementation, and I encourage anyone interested to both read this work, and determine whether they prefer the goal of equitable advertisement to be left to advertisers and users or to be enforced at the platform level.
How much revenue might advertising platforms lose with this approach?
A quick analysis suggests the revenue lost in thinning the market
this way should be quite manageable, and possibly nonexistent,
depending upon what targeting high-stakes advertisers were allowed in
the unified market.
Suppose a platform has 5% of its advertisers who place housing advertisements. These advertisers contribute to the revenue of the platform in two ways. First, they purchase some ad slots from the company. Second, even for advertisements they ultimately do not buy, they compete with other advertisers, driving up the price of those advertisements.
How much revenue must be lost to each of these sources?
For those impressions that housing advertisers won in the unified market, revenue can be lost where non-housing advertisers provided competition and drove up the housing advertisement prices. If this competition was untargeted, then this can be replaced with effective setting of reserve prices, resulting in little or no revenue loss. If the competition was targeted to 10% of the population, this can still be recovered by randomly setting reserve prices on 10% of the housing ad slots. The only revenue that cannot effectively be recovered using reserves comes from settings where some housing advertiser was targeting and willing to pay more for certain populations than others. So, some revenue is lost if the comparison is done to a system that allows housing advertisers to target, but none is lost compared to one where housing ads cannot target.
For ads that housing advertisers did not buy but drove up prices, a similar
analysis can be done — cleverly set reserve prices can “mimic” the competition the housing advertiser provided. This could even be done in a (demographically) targeted way, since the advertisements in this segment are only advertising target-appropriate content. So, the ads in the targeted segment of the market can generate at least as much revenue as they did in the unified market.
What other concerns does the proposed “segmented” market for advertisements raise?
If we shift from a completely unified market for advertisements to one where 10% of all advertisments are preordained as “un-targetable” impressions for high-stakes ads, what additional impacts might this have? The primary one, from my perspective, is that it will lead to necessary decisions about which ads are un-targetable and which are targetable. Who will make this determination?
If the advertisers self-identify, they will do so in a way to maximize their utility. This may or may not lead them to truthfully report whether their advertisement pertains to, say, housing. If it is sufficiently profitable for them to target, they will report their ad belongs to a targetable segment; if it is profitable enough for them to not target but pay possibly lower prices in the untargeted market, they will prefer that. Equitable display of untargeted categories will not be affected by targetable ads behaving as un-targetable, though it will reduce the rate at which anyone sees un-targetable categories. On the other hand, un-targetable ads such as housing ads might have considerable incentive to target, and in targeting may very well create inequity in the display of housing advertisements. Advertisers who categorize their own ads as targetable, then, will need to be audited by either the platform or some external system to ensure such ads do not pertain to some un-targetable category.
[1] Datta, Amit, Michael Carl Tschantz, and Anupam Datta. “Automated experiments on ad privacy settings: A tale of opacity, choice, and discrimination.” arXiv preprint arXiv:1408.6491 (2014).
[2] Sweeney, Latanya. “Discrimination in online ad delivery.” Communications of the ACM 56.5 (2013): 44–54.
[3] https://www.propublica.org/article/facebook-ads-age-discrimination-targeting
[4] https://nationalfairhousing.org/facebook-complaint/
[6] https://www.propublica.org/article/facebook-lets-advertisers-exclude-users-by-race
[7] Nasr, Milad, and Michael Carl Tschantz. “Bidding strategies with gender nondiscrimination constraints for online ad auctions.” Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency. 2020.
[8] Ilvento, Christina, Meena Jagadeesan, and Shuchi Chawla. “Multi-category fairness in sponsored search auctions.” Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency. 2020.
[9] Chawla, Shuchi, and Meena Jagadeesan. “Fairness in ad auctions through inverse proportionality.” arXiv preprint arXiv:2003.13966 (2020).