Discrimination through Image Selection by Job Advertisers on Facebook

Facebook’s ad system consists of two phases - ad creation and ad delivery (Facebook, 2020b, 2023a). During ad creation the advertiser chooses (1) their business objective, i.e. an outcome they’d like to achieve, such as increasing the number of visitors to their website, (2) contents of the ad including the images, text, and destination page link, (3) the targeting parameters, specifying the kinds of users they’d like the ad to be shown to, and (4) the budget and duration of their ad campaign. Advanced targeting capabilities, using demographic, geographic, interest and/or behavioral characteristics of users have long been touted by Facebook as a useful tool for advertisers.

Ad delivery is the process through which a subset of the users targeted by the advertiser are chosen to see the ad. For each user, the platform runs an auction among all advertisers targeting that user, to determine which ad to show. The high-level information publicly known about the auction is that it does not necessarily select the ad with the highest bid; rather, for each candidate ad, Facebook computes its Total Value, defined as:
$\textit{Total Value = }\textit{Advertiser Bid}\times\textit{Estimated Action Rate}+\textit{Ad Quality};$ and then shows the user the ad with the highest total value (Meta, 2023b; Facebook, 2020b, 2023a). All components of this equation are computed using machine learning. The Estimated Action Rate is Facebook’s predicted likelihood of the user taking an advertiser’s desired action, the Ad Quality is a machine learning model that combines feedback from users and predictions based on image and text of the ad creative, and even the Advertiser Bid is an algorithmic model, where Facebook chooses what to bid on behalf of the advertiser based on their budget and expressed preference in bid strategy (Facebook, 2022a).

Prior work has found significant evidence of Facebook’s extensive targeting tools being used in order to exclude individuals on the basis of gender, age, or race in housing and employment advertising (which are areas governed by U.S. anti-discrimination laws). For example, Facebook tools offered the ability to exclude Black and Hispanic users living in a specific area from the targeting of housing ads (J. Angwin and T. Parris Jr., 2016; Tobin, 2019). Furthermore, (Tobin and Merrill, 2018) found ads from ten traditionally male-dominant industries targeted just at men, including a software company, a moving company and a police department. Moreover, (Angwin et al., 2017) found that advertisers excluded older workers from seeing job ads. For instance, financial analyst job ads were only targeted at users aged 25–36 years.

Discrimination can occur not only due to advertisers’ choices during targeting, but also as a result of the ad delivery process, as the machine learning driven components of the total value equation can be biased.

Specifically, Ali et al. (2019) showed that in the case of Facebook, the ad creative image influences and skews the ad delivery along gender and racial lines, in ways that cannot be explained by market effects or users’ interactions with the ads. Follow-up work by Imana et al. (2021) demonstrated that the skew in ad delivery by gender in the case of employment ads cannot be justified by differences in qualifications among demographics in the target audience, and is thus fully attributable to algorithmic decisions made by Facebook in its own business interests. More recently, concurrent and complementary work by Kaplan et al. (2022) showed that ad delivery can be dramatically skewed solely as a result of demographic characteristics of the people depicted in the images. These works provide evidence of discrimination in ad delivery, a hypothesis that was put forward by Sweeney (2013) and further investigated by (Datta et al., 2015, 2018; Lambrecht and Tucker, 2019; Kingsley et al., 2020; Celis et al., 2019; Dwork and Ilvento, 2018).

In response to the investigative reporting (J. Angwin and T. Parris Jr., 2016; Tobin and Merrill, 2018; Angwin et al., 2017), lawsuits (Tobin, 2019; ACLU, 2018a), and a Civil Rights Audit (Facebook, 2020a), Facebook disabled the advertiser ability to explicitly target employment ads by age or gender in 2019 (Sandberg, 2019). The ability to explicitly target by race and ethnicity had been removed earlier in 2018 (Facebook, 2018); however, proxies for targeting by racial categories remained available even in 2022 (Keegan, 2021; Facebook, 2022b). Facebook also introduced a requirement for advertisers to self-identify when running ads on housing, employment, or credit (HEC) issues, and to acknowledge adherence to Facebook’s Discriminatory Practices policy (Facebook, 2019, 2023b). Google (Google, 2020) followed suit by restricting targeting criteria for HEC ads and requiring advertisers to acknowledge adherence to a personalized advertising policy. There is no advertiser-facing feature on Google to self-identify the ad category; instead, ads may be automatically labeled as belonging to an HEC category after creation and review. LinkedIn retained the ability for advertisers to target based on age or gender, but required advertisers to self-certify that for their HEC or education ads, they will not use the platform to discriminate based on age or gender (LinkedIn, 2021).

Addressing discrimination in ad delivery remains more complex. In a June 2022 settlement with the Department of Justice, Facebook committed to “develop a new system to address racial and other disparities caused by its use of personalization algorithms in its ad delivery system”, called the Variance Reduction System (VRS) (The United States Department of Justice, 2022). Specifically, the VRS system, aims to ensure that the age, gender and estimated race distribution of the audience that is shown a housing ad closely resembles the distribution of age, gender and estimated race of the audience targeted by the advertiser (Meta, 2023b). During an ad’s delivery, Facebook periodically compares the ratio of impressions delivered to a particular age / gender / racial subgroup, with that subgroup’s fraction in the advertiser’s target audience. When the ratios diverge, Facebook adjusts one of the machine learning algorithms used to calculate the ad’s total value in the auction, in a way that will change the likelihood that this ad will win the auction and be shown to a user of a particular subgroup. Although as of January 2023, the VRS system is implemented only for housing ads, Facebook has committed to expanding it also to employment advertising (Austin Jr, 2023).

A limitation of the VRS approach for eliminating discrimination is that the demographic composition of the advertiser’s targeted audience is chosen as the goal for the composition of the delivery audience. Such a choice may be problematic as long as the advertisers continue to have access to tools (e.g. custom audiences, location based audience creation) to select their targeted audience in a discriminatory manner (Speicher et al., 2018; Sapiezynski et al., 2022; Faizullabhoy and Korolova, 2018).

As will become clear from subsequent discussion in Sections 2.3.1 and 2.3.2, the VRS approach is unlikely to address the new type of discrimination through image selection that we identify in this work. Extrapolating from the social science literature and case law (Walker et al., 2012; White et al., 2019; Thaler-Carter, 2001; Avery and McKay, 2006; Johnson, 1991), the selective use of people in ad images affects the ad recipients’ likelihood of acting on the ad (e.g., their likelihood of clicking on it), whereas the VRS is entirely focused on minimizing the variance in impressions (i.e. showings of the ad).

The Facebook Ad Library²²2https://www.facebook.com/ads/library/ is a transparency tool created by Facebook that provides a keyword-based interface to a searchable collection of active ads on the platform. After choosing an ad category among (1) all ads, (2) housing, (3) employment, (4) credit, or (4) issues, elections or politics (referred to as political ads henceforth), one can search for ads based on specific keywords or suggested advertiser names.

The search results page loads ads in the category chosen in reverse chronological order as one scrolls down the page. For all active ads, the search results page contains details about the ad content used, including one or more of the ad text, destination page link, video, and image(s), and other meta data such as the Ad ID, and the date it started running. However, it does not contain data about inactive ads. Critically, with the exception of political ads, neither the ad targeting information (such as advertiser targeting choices or campaign budget) nor the ad delivery information (such as the number of people an ad reached, location where it was delivered, or a demographic breakdown of the ad recipients) is available.

Most prior works using the Facebook Ad Library focused on political and issue advertising (Edelson et al., 2020; Le Pochat et al., 2022; Silva et al., 2020; Sosnovik and Goga, 2021). We are aware of only one effort that uses the data in the ad library for inferences about employment advertising — a class action charge by Real Women in Trucking against Meta (formerly Facebook) alleging that Meta discriminates against women and older people when deciding which users receive employers’ job ads on Facebook (Romer-Friedman and Ebadolahi, 2022). In it, the supporting evidence lists over 75 employment ads from the Ad Library that were delivered primarily to men. For example, an employer seeking truck drivers in North Carolina reached an audience that was only 5% female.

Images play a unique role in advertising due to their distinctive characteristics, which are not present in other sources of media (Messaris, 1997; Smith, 2008). In his 1997 book “Visual Persuasion: The Role of Images in Advertising” (Messaris, 1997), Paul Messaris outlines iconicity, indexicality, and syntactic indeterminacy as the fundamental traits of images that distinguish them from other modalities and uniquely affect the ad campaign. First, the iconicity of images gives advertisers access to strong emotional responses invoked in the target audience when they view an image. Second, images are indexical and can serve as documentary evidence of an advertiser’s point of view. Third, images lack explicitness and syntax, and thus the audience’s interpretation of the message conveyed through an image in an ad campaign is a creation of their own.

Based on social science theories, researchers argue that images present on recruitment websites and in advertising affect the application intention of job seekers and can be used to manipulate the gender and racial composition of those who apply. Recent works (Walker et al., 2012; White et al., 2019) applied Spence’s Signaling Theory (Spence, 1973), Social Identity Theory (Tajfel and Turner, 1974) and Visual Perception Theory (Gibson, 2002) in support of this hypothesis. Based on the Signaling and Perception Theories, (White et al., 2019) argued that people who apply for jobs will resemble in their demographic characteristics those of the people depicted on the job websites. In an apparent attempt to leverage the effects of these theories in practice, organizations have been found to use images of diverse people in job ads to recruit diverse talent (Thaler-Carter, 2001; Avery and McKay, 2006). In parallel, (Walker et al., 2012) leveraged the Social Identity Theory to deduce that “individuals naturally categorize themselves and others in terms of important visible characteristics (e.g., race, gender, age)”, and “develop more favorable attitudes toward in-group members and seek environments that affirm their identity”. As a result, they demonstrate that job seekers infer diversity cues from the people present in the images, which in turn influences their job application decision.

Legal scholarship under both Section 704(b) of Title VII (U.S.C., 1964) and the parallel, but broader Fair Housing Act (FHA) (U.S.C., 1968) of U.S. law suggests that use of images containing only people of the dominant demographic in jobs or housing ads may be interpreted by the casual observer to convey a preference based on a protected class, irrespective of the advertiser’s intent (Capaci v. Katz & Besthoff, Inc., 1983; Datta et al., 2018). Evidence of such discrimination is detailed in the FHA case law (Johnson, 1991). According to (Datta et al., 2018), “the FHA case law connects the ordinary reader standard to the prohibition on sex-designated advertising columns (e.g., those with “Male Help Wanted” and “Female Help Wanted” headings) by explaining that advertisements that exclusively feature White models may discourage Black people from pursuing housing opportunities by conveying a racial message in much the same way that the sex-designated columns furthered illegal employment discrimination” (see Section 4.1.2 for additional discussion).

FHA case law can reasonably be extended to apply to the context of employment ads as evidenced by a high profile settlement involving Abercrombie & Finch in 2004 (Illston, 2004). As part of that settlement, Abercrombie agreed to add more diversity to their marketing materials and reflect the makeup of the nation’s population; so as not to discourage members of minorities from applying for jobs that were otherwise dominated by “well-known collegiate, all-American - and largely White” people (Greenshouse, 2004).

The influence of images in shaping biased perceptions or exaggerating diversity has been extensively researched across various contexts beyond online advertising. Moriearty (2009) argued that the over-representation of Black youth as criminals in images used by media has resulted in disproportionately stricter policing and harsher sentencing for Black youth. Similarly, in an effort to attract under-represented members and create a misguided sense of belonging, universities have been found to over-represent African American students in their admissions brochures (Pippert et al., 2013). Netflix too has been accused of using misleading visual representations to entice viewers based on race; for example, some marketing posters contained only Black people even if they appeared only briefly in a show (Zarum, 2018). In fact, a 12 year longitudinal study of TV shows by Baruah et al. (2022) found that male characters with a light skin tone occupy the majority of the screen time. Biases have also been identified in reports on vaccination and antimicrobial resistance; with White people shown in a professional capacity whereas children of color shown as vulnerable and exposed (Charani et al., 2023). In the Computer Science community, the works of (Metaxa et al., 2021; Papakyriakopoulos and Mboya, 2021; Kay et al., 2015) found a gender- and race- driven representation bias in both image- and text-based Google Image Search results. Similarly, Luccioni et al. (2023) found that Text-to-Image models depict care professions (e.g., dental assistant) nearly exclusively through as women, and positions of authority (e.g., director, CEO) – exclusively through men.

To quantify how prevalent is the selective use of gendered people in job ad images, in January 2023 we scraped data from the Facebook Ad Library and saved the job ad images for two occupations - truck driver and nurses. We restricted our search to the employment ads category and to only those ads which contain images and memes. We chose these occupations because (1) they currently have a real world defacto skew towards a specific gender (91.9% of truck drivers in the US are men, 87.9% of nurses in the US are women (Bureau of Labor Statistics, 2021)) and (2) we found thousands of ads for these occupations in the Ad Library, allowing for a large scale analysis.

Since the Ad Library does not load all search results when there are more than a few hundred, obtaining data on all advertisers running employment ads for truck drivers or nurses is not merely a task of performing one search. Instead, to stay within the number of results that load, we turned each keyword search into 26 separate searches by expanding a given keyword with an additional letter of the alphabet, i.e. we searched and downloaded results for the keywords truck driver a through truck driver z and nurses a through nurses z³³3A sample URL for the truck driver a keyword: https://www.facebook.com/ads/library/?active_status=all&ad_type=employment_ads&country=US&q=truck%20driver%20a&search_type=keyword_unordered&media_type=image_and_meme. For each job advertiser whose ads appeared in the search results, we recorded their Page ID, and then performed a separate search⁴⁴4A sample URL for Truck Driver Recruiting America: https://www.facebook.com/ads/library/?active_status=all&ad_type=employment_ads&country=US&view_all_page_id=106206904474383&search_type=page&media_type=image_and_meme to download all the images, excluding the videos, of all the advertiser’s active ad campaigns at the time. We explored different variations of the keywords such as singular or plural forms and whether or not to include quotes⁵⁵5A search for the keyword enclosed in double quotes returns results with an exact match in the content of the ad., and ultimately settled on the form that produced the highest number of search results⁶⁶6See Appendix A.1 for precise queries and select advertiser URLs.. Although our data collection may have excluded some advertisers (e.g., those whose ad text contains only the words nurses or nurses followed by digits or special characters), we believe our sample is large enough to draw conclusions.

Since an advertiser may re-use the same image in many campaigns, for each advertiser, we considered both the total number of images used and the number of distinct images used. We determined whether two ad campaign images are distinct using a hashing algorithm of the open source imagededup tool (Jain et al., 2019).

We annotated each image for whether it contains people and if so, the perceived gender of the people in it. We did not consider people in the images who were not clearly representatives of the occupation being recruited for (e.g., children who the nurse was attending to, or family of a truck driver waiting for them at home). Further, we corroborated our results by recruiting crowdworkers through Amazon Mechanical Turk (MTurk) who were paid above the federal minimum wage. As a result, each distinct image was annotated by 2 distinct annotators. We excluded from analyses the advertisers for whom we could not reach an agreement on annotation, or were unable to decisively determine the gender of all people due to occlusion, blurriness or use of avatars. We summarize our annotated data in Table 1.

We find that discrimination through image selection is a prevalent practice among the advertisers we study. Specifically, among 159 truck driver advertisers whose campaigns contain people, 90 or 57% depict only men as truck drivers. Among 259 nurses advertisers whose campaigns contain people, 108 or 42% depict only women as nurses. Although we observe that the fraction of advertisers using images depicting only people of the stereotypical gender for their profession typically decreases as the number of distinct images used goes up; we find examples of extensive stereotypical image selection across all ranges of the number of images per advertiser. Specifically, among the 33 truck driver advertisers who use only 1 image, 25 or 76% depict a man; whereas among the 28 advertisers who use 6-9 distinct images, 13 or 46% depict only men. For the 29 advertisers in our dataset each using at least 15 distinct images, only 4 (but still four and not zero), or 14% depict only men. See Table 2 for the detailed statistics broken down by the number of distinct images used and Figure 3 for examples of stereotypical image selection by two advertisers.

Figure 4 compares the rate of representation by gender (women) and race (White and Black) of people used in job ad images to the rate for that demographic characteristic per corresponding BLS occupation for each advertiser. Specifically, for a particular demographic characteristic and advertiser, the X-axis represents the fraction of people with that characteristic for that advertiser’s occupation per BLS, and the Y-axis represents the fraction of people in the job ad images. We plot a 45-degree reference line to visually assess the magnitude of over- or under-representation given by the regions above and below the reference line, respectively. Full advertiser specific statistics are presented in Table 4 in Appendix A.2.

Across the 15 advertisers in Figure 4 and columns 6-8 of Table 4, we observe that most advertisers over-represent women and Black people, and under-represent White people. We draw definitive conclusions for advertisers for whom the error bars do not cross the reference line, finding the following:
Women: 10 (over), 4 (under), 1 (inconclusive);
White People: 0 (over), 12 (under), 3 (inconclusive);
Black People: 8 (over), 1 (under), 6 (inconclusive).

Our findings indicate that advertiser selection of people in job ad images with regards to gender and race is complex and varies across different occupations. Broadly, across a diverse set of occupations we find that some advertisers (e.g., NYPD Recruit and TSA) may be attempting to diversify their existing workforce by including more women and Black people than are employed in those professions per BLS in their images.

The questions posed by our study and the challenges encountered while performing it imply that, at a minimum, the Ad Library for employment ads should provide the same data as is already being provided for social issues, elections or politics ads. Specifically, it should enable obtaining a list of all employment advertisers (rather than require scraping workarounds we had to resort to in Appendix A.1). Then for each advertiser, it should provide information on the targeting criteria and budget chosen by the advertiser for each of their ads, as well as delivery information – the number of people shown the ad, broken down by their demographic characteristics. To study changes in advertiser image selection practices over time, the Ad Library should include data for both active and past campaigns. Specific to considerations for discrimination in employment targeting, the Ad Library should provide an easy way to download the images used by the advertisers, label (or ask advertisers to self-identify) the job industry of the ad, and provide aggregate information about the people who engaged with the ad, broken down by their demographic characteristics. Furthermore, rather that aggregating the targeting data at an advertiser level or on a weekly basis as is done for political ads, employment ad data should be provided at an ad campaign level over the duration of that campaign. This desiderata echoes the one called for by (Edelson et al., 2021; Ali et al., 2021; Imana et al., 2021), and is feasible given the existing implementation for political ads.

However, as is clear from the works related to discrimination due to ad delivery optimization, discussed in Sections 2.1.3 and 3.4, transparency of advertiser image selection practices is not sufficient to understand the implications of such choices, neither for public-interest researchers nor for the advertisers themselves. We thus call for either disabling of the use of ad delivery optimization (and the opaque machine learning algorithms underlying the total value equation) for employment advertising or a significant increase in transparency of these algorithms when applied to job ads. A viable path for this kind of transparency has recently been proposed by Imana et al. (2023). Specifically, their proposal is a platform-supported special access API that gives public-interest researchers an ability to query the values of the estimators used in the Total Value equation, and study how they vary depending on the perceived gender and race of the person in the image and the platform-inferred demographic of the potential ad recipient. The approach of (Imana et al., 2023) addresses the concerns raised by platforms in response to requests for greater accountability – those of user privacy and protections of the platforms’ proprietary code and business interests – through the use of differential privacy and the specifics of the set-up of the auditor - platform API interaction.

The platforms should take proactive action towards ensuring that advertisers do not deliberately discriminate through their image selection. In fact, they already do so in a somewhat analogous domain of potentially discriminatory text selection by advertisers.

Similar to the case of selective image choice influencing who applies for jobs, prior social science work shows that selective use of language affects ad recipients’ actions (Gaucher et al., 2011; Sraders, 2019) (e.g. removal of the word “aggressive” from the ad text increases the number of women applicants). Legal precedent establishes that the practice of printing classified ads in two separate columns: “Help wanted: Male” and “Help wanted: Female” violates anti-discrimination law (National Organization For Women (2014), NOW; ACLU, 2018b; Kerman, 1974; New York Times, 1973). Thanks to the precedent and a relative ease for identifying discriminatory language use compared to discriminatory image use, advertising platforms such as Facebook (Facebook, 2023c) and LinkedIn (LinkedIn Corporate Communications, 2019) already have advertiser guidance about the language for creating inclusive ads, and may be enforcing them at ad review time. We suggest platforms extend the guidance to image use, e.g., by updating their policies and advertiser training materials.

Furthermore, as platforms themselves gear up to deploy Generative AI-based tools to assist advertisers in their ad campaign and creative creations (Okudaira, 2023; Meta, 2023a), they should be mindful not to lead the advertisers to employ image selection in a discriminatory manner.

One approach to consider is through the adaptation of practices for ensuring fair and impartial jury to the advertising context. In the U.S. that means that a pool from which juries are selected should satisfy a fair cross section requirement (Seltzer et al., 1995; Gastwirth et al., 2015) via a Duren Test (Duren v. Missouri (1979), U.S. Supreme Court). It measures the extent to which the jury pool demographics differ from those of the community, where the relevant community consists of individuals who are eligible for jury service. Several disparity metrics are used to compute the differences in demographics, and case law establishes permissible levels of disparity (see (National Center for State Courts, 2010) for an overview). The challenge of adopting it in the advertising context is that the equivalents of “eligible for the jury service” in the “relevant community” is difficult to establish – both who is qualified for the job and who constitutes the relevant community for a particular job (e.g., due to commute limitations) are open questions. In the jury selection case, this data is relatively easily available through the U.S. Census. Prior work in the computer science literature (Metaxa et al., 2021) adopts the approach of comparing the racial and gender composition of Google Image Search results to that of the composition in the U.S. workforce across specific occupations.

Literature on ensuring compliance with the Individuals with Disabilities Education Act, and determining whether disproportionalities on the basis of race exist in the placement or discipline of students with disabilities at the school district level may also be of inspiration (Farkas et al., 2020; U.S. Department of Education, 2016). However, it also operates in the context of full information, basing calculations on the exact number of students of particular race in a particular school district that is known – an assumption that does not hold for the advertising context.

Additional metrics may need to be established to assess the image use of advertisers who may be trying to affirmatively reach underserved populations or attempting to diversify beyond the demographics of the existing U.S. workforce for specific occupations. For example, the primary goal of Black Career Network (Page ID 70203958046) and Black Excellence KC (Page ID 563610990802939) is to recruit Black professionals. Therefore they could be justified in selecting many (or only) Black people for their job ad images. Complementarily, in our study we found examples of advertisers attempting to diversify their workforce beyond the existing U.S. workforce of specific occupations by over-representation of women or Black people, or under-representation of White people in their job ad images.

Accounting for such practices is difficult: Is the choice of advertisers to empower an under-represented group through their ad images ethical or is it legal? Can this tension be reconciled if the advertiser self-identifies their reasons for doing so? How long should such practices be permitted? Objections have been raised to such practices in the past, e.g., White males have brought lawsuits for “reverse discrimination” against employers who have taken affirmative action to ameliorate the impacts of past discrimination (U.S. Equal Employment Opportunity Commission, 1981). The approach for accounting for such practices remains an open question, represents an ongoing policy issue (Austin Jr, 2023) and is presently the subject of active judicial scrutiny.

Ultimately, since the advertiser’s image choice is only one of the inputs to an extremely complex and opaque system, one can argue that the desiderata should be placed on the eventual outcome of ad delivery, rather than on the targeting, and the burden of flagging potential discrimination be shared between the advertiser, the platform, and public-interest researchers. We believe the exploration of such approaches is an important area for future work.