The Algorithmic Audience and African American Media Cultures
Tim Havens / University of Iowa
Audience measurement has been a longstanding (if not terribly sexy) issue in African American media studies. For decades, audience numbers were reported as Gross Ratings Point (GRPs), or the aggregate percentage of metered homes that watched a particular broadcast. As early as the 1977 report by the U.S. Commission on Civil Rights, entitled “Window Dressing on the Set: Women and Minorities in Television,” observers began to point out that the broadcasting industry’s reliance on GRPs made it tough for anything that is non-mainstream, non-hegemonic, non-supremacist, and non-patriarchal to find its way on air. In the fallout of this report, Nielsen began over-representing black households as a percentage of their panels in order to move the needle at least a little bit when black and white tastes diverged substantially.
As Ien Ang has argued, no audience research ever measures real viewers’ and their tastes.  Instead, they construct a particular picture of audience that is deeply influenced by the technologies used to measure them. When measurement techniques change, Eileen Meehan has shown, so does the dominant image of the audience.  This dominant image, moreover, circulates among programming executives who make production and acquisition decisions based upon that image.
By the time the Commission on Civil Rights’ report was published, the networks had already begun to integrate demographic considerations into their understanding of the audience, thanks to advances in Nielsen’s audience measurements. As the 1980s and 1990s progressed and the networks began to shed viewers to cable channels, they began to focus much more on their core 18-49 year old white family audience; since then, as Herman Gray argues, the networks have considered African Americans as political subjects capable of causing political turmoil for the networks, but not as economic subject worth targeting consistently with relevant programming. This set-up created a predictable dynamic between African Americans and the networks, as the networks inevitably dropped black-oriented shows for poor overall ratings, followed by political agitation on the part of African American and other minority-based political groups, which led to a brief surge in minority programming. 
Today, we have seen another revolution in audience measurement with the explosion of digital data and the development of data-mining algorithms that make sense of viewers, tastes and behaviors in new ways. While the industry long lived in an era of scarcity of audience data, today there is an overabundance.
How might these new forms of algorithmic audience measurement shape the media culture we inhabit? I have primarily begun to think about this question through the streaming service Netflix. Netflix exhibits an odd contradiction: it exhibits a range of programming about (and sometimes by) African Americans and other minority groups, including Dear White People, Orange is the New Black, and Narcos, but it also has a reputation among some subscribers and independent producer as insensitive and closed to minority tastes and content producers. Among some African American subscribers, it has become a commonplace that, once they watch a single black-cast television series or film, they are suddenly inundated with every other black-cast offering on Netflix. Seemingly, the algorithm thinks that black people are only interested in black-cast content, and that everyone who watches a black-cast film or TV series must be black. There’s even a sentiment that circulates among some African Americans that Netflix’s black-cast offerings, as compared to their predominantly white-case offerings, are inferior in quality and steeped in stereotypes.
The idea that Netflix is largely insensitive to African American tastes is only one perspective, and it may well be a minority one at that. Still, at a time when the fate of cultural diversity on screen is in the hands of algorithms, the people who program them, and the people who interpret their findings, it is worth asking how they are shaping the diversity of the media content available through streaming services.
Here, I sketch out a typology of how to study the role of algorithmic audience analysis in commercial African American streaming culture, including questions of recommendations and user interface, content availability, and programming decisions. What results is a sort of research agenda, parts of which are certainly much easier to research than others.
Racial bias and exclusion in recommendation algorithms can happen at different moments in the process. At the input moment, it’s possible that African American are absent (or nearly absent) from the universe of subscribers in the first place. Indeed, Horowitz Research, who in 2015 published a report titled State of Cable & Digital Media: Multicultural Edition, found that African Americans tend to watch more television than other ethnic and racial groups, just as countless other research studies have shown for decades. In addition, they found that African Americans living in urban areas oversubscribe to premium television services, compared with other urban ethnic/racial groups. However, Horowitz also found that African Americans undersubscribe to Netflix, even as they purchase more pay-per-view programming and oversubscribe to Hulu: while 57 percent of all urban viewers subscribe to Netflix, only 56 percent of African Americans do. Granted, the difference is small, but it’s three percent less than white urban viewers, and in every other category of programming, percentages of African American viewers exceed those of white and other ethnic groups. In other words, African American under-subscription to Netflix certainly stands out in the report.
Why do African Americans subscribe to Netflix at lower rates than other groups? This of course is a much tougher question. One might suspect that broadband internet penetration rates might be to blame: for cost reasons, African American broadband penetration rates do tend to lag a few percentages points behind most other racial and ethnic rates. However, if that were the sole cause, we would also expect undersubscription to Hulu, which isn’t the case. Instead, there may be content issues with regards to Netflix that explain African American subscription rates as well.
It may also be the case that the Netflix library offers little of interest for African American viewers, driving undersubscription rates because those potential subscribers know there will be little content for them. Content bias is of course the linchpin of the question of racial bias in Netflix’s algorithm, since relevant programming is at the heart of longstanding concerns about race and media. Empirically, this is a difficult topic to study, given the vastness of the Netflix library and the company’s licensing arrangements with content owners, which can cause programming to come and go from the library quite frequently. Finally, no good quantitative measure of what might constitute “programming for African American subscribers” has ever been developed, nor could it ever be.
If content diversity is a difficult object to fix methodologically, it is less difficult to imagine how we might study how Netflix executives use algorithmic data to make programming decisions. However, given the proprietary nature of algorithmic audience data and Netflix’s tight-lipped approach to releasing data and discussing content acquisition decisions makes addressing this question directly thorny as well. In the absence of such information, we can rely on some extant data that suggests that Netflix’s original programming, at least, is probably not designed with African Americans primarily in mind. According to the 2015 Hollywood Diversity Report, streaming television series creators, directors, and writers (the vast majority of whom, at the time, must have been working on original Netflix series) are substantially whiter than their counterparts in cable or broadcast television. Of course, original programming is only a fraction of Netflix’s content, and may not be a main factor for deciding to subscribe to the service. Nevertheless, since original programming is both signature content and a loss leader for streaming television, the fact that such series seem to be designed mainly for white audiences lends credence to the impression that Netflix’s overall content acquisition practices may privilege white subscribers as well.
By way of closing, I want to talk a little about some new research I’ve been working on with an interdisciplinary team of humanists, social scientists, and computer scientists. We have been looking at the racial/ethnic discrimination in recommendation and filtering algorithms. Recently, we have started examining the question of input bias in streaming media service by interviewing subscribers and non-subscribers of various races and ethnicities about why they choose to subscribe to Netflix or not, and what their experience with Netflix’s offerings and recommendations has been. In addition, we are planning more systematic probing of how differential input into the Netflix system results in differential output, and whether we can find any pattern or logic in the personalized results. We have used a similar design to show that Google News personalizes search results based upon a user’s social media activity, but the complexity and variability of the Netflix user interface will create a good deal of unexpected problems.
This remains very preliminary research, and it requires a substantial amount of time, a range of expertise, and the development of new research methodologies. At each moment in the streaming media process – the input, the algorithmic processing, and the interpretation of the data – we have very little reliable information available, and we need to be creative and collaborative if we hope to find good ways to get more. Still, if the main questions that have animated African American media studies since the days of broadcasting are going to continue to concern us in an era of streaming, we will need to develop these new tools and modes of scholarship. It is a big job.
Please feel free to comment.
1. Banner image