Abstract
This essay asks two related questions: what is unique about streaming services (and Netflix specifically), that generates a greater investment in the diversity of its content, and how does the technology associated with streaming, in particular algorithmic recommendation systems, facilitate an engagement with diversity and inclusion? To answer these questions the essay considers the relationship between Netflix’s Inclusion Strategy, its Recommender Algorithm, and the diversity of its content, exploring the complex set of relations that exist between the way Netflix recommends content to its audience and its (perceived) diversity.
Introduction
The zeitgeist shift signaled by social movements such as #MeToo, #5050x2020, #TimesUp, and #OscarsSoWhite marks an acknowledgment by the film and television industries that they need to do better in terms of improving the representation of, and respect toward, marginalized groups in front of and behind the camera. In this landscape, streaming services are arguably leading developments in the industry in this area. 1 In January 2021, Netflix released the first report of its newly launched Inclusion Strategy, examining all films and series commissioned by the company between 2018 and 2019 against twenty-two indices of diversity and inclusion, including sexuality, disability, race, and gender (Myers 2021). It showed many positive gains over this two-year period in relation to the diversity of on-screen talent as well as creators, producers, writers, directors, and cinematographers. The Netflix Inclusion report is acknowledged by the company as being U.S.-centric, with race and ethnicity data only collected from Netflix’s U.S. employees, although it is noted that women make up 51.7 percent of Netflix’s global workforce (Myers 2022).
As I write this essay from an Australian context, primarily with access to the Australian Netflix catalog, the debate around diversity and inclusion in the Australian screen industries is also growing. Screen Australia’s recent report, Seeing Ourselves: Reflections on Diversity in TV drama (2016) notes that the majority of characters we see on Australian television continue to be straight, able-bodied, and Anglo-Celtic, despite the increasing diversity of our population and audiences (Australian Bureau of Statistics 2016). Yet, as the report also observes, there is a “new momentum and appetite for change across the industry,” particularly in relation to issues of inclusion and diversity.
Netflix is a global streaming service, and diversity matters. But diversity also means different things regionally. It involves telling stories that reflect the local population, although from the perspective of a global company like Netflix these stories must also cross borders. After all, the Netflix Inclusion Strategy is not only about diversity; it is also about reach. The unexpected popularity of the South Korean series Squid Games on Netflix in 2021 demonstrated how a niche Korean-language series can become a global hit: Squid Games became the most highly viewed Netflix series in 94 countries, surpassing Bridgerton, a period drama with a racially diverse cast, as the most watched Netflix series of all time (Spangler 2021). Where traditional television is more conservative in its programming to meet the demands of advertisers, subscription-based services such as Netflix are thriving by offering diversity, or at least the semblance of diversity, to its audiences.
This essay seeks to address two related questions: what is unique about streaming services (and Netflix specifically), that generates a greater investment in the diversity of its content, and how does the technology associated with streaming, in particular algorithmic recommendation systems, facilitate an engagement with diversity and inclusion? To put this another way, are streaming services presenting more diverse stories than traditional television services or are they simply better at marketing and promoting to niche audiences? To answer these questions, I consider the relationship between Netflix’s Inclusion Strategy, its Recommender Algorithm, and the diversity of its content, exploring the complex set of relations that exist between the way Netflix recommends content to its audience and its (perceived) diversity.
The essay proceeds with a discussion of the Netflix Recommender Algorithm, a collective term for a series of proprietary computational tools developed by the company since the early 2000s. As a subscription-based service, Netflix relies on a loyal subscriber basis, coupled with a need to grow this membership, and carry a broad catalog that will appeal to this audience. In this regard, personalization systems and algorithmic filtering are extremely important to how Netflix curates its content for its subscribers. The essay then goes on to explore Netflix’s use of a new algorithm in 2017 to deliver personalized artwork to its subscribers, providing, at least on the surface, more diversity of content by imagining (and literally imaging) its catalog in multiple and alternate ways.
Methodologically, the essay combines an analysis of articles from Netflix’s technical and research blogs, and company media releases, with news reports and social media posts about Netflix algorithms, particularly instances where encounters with these algorithms “are brought to the fore during breakdowns, accidents, and controversies” (Gaw 2021, 7). I provide a semiotic analysis of specific instances of artwork personalization to show how recommender systems cater to different taste communities and, in doing so, appeal to a diversity of audiences.
The Netflix Recommender Algorithm (NRA)
Recommendation systems are “algorithmic tools that internet platforms use to identify and recommend content, products, and services that may be of interest to their users” (Singh 2020, 6). Algorithms themselves can be defined as “a set of instructions, rules, and calculations designed to solve problems” (Benjamin 2019, 11). Netflix’s “problem” is how to appeal to and diversify its audience so as to retain and increase its subscriber base. Unlike other major streaming services, Netflix’s financial success relies solely on its ability to attract and retain subscribers. Of its main competitors, Apple TV channels users into larger media ecosystems, and Amazon Prime Video forms part of a broader e-commerce entity. Other streaming sites such as YouTube rely on targeted advertising for their revenue. Netflix, however, has a major incentive to retain and grow its stable of subscribers. So valuable are its recommendation systems that in October 2006 the company launched the Netflix Prize, a contest offering US$1m to the first team to develop a recommendation system capable of predicting movie ratings with at least 10 percent greater accuracy than its existing system, Cinematch. Concluding on 21 September 2009, the competition drew more than fifty thousand participants from 186 countries (Hallinan and Striphas 2016). 2
Central to Netflix’s brand and business is the Netflix Recommender System (NRS), “a collection of proprietary algorithms used to recommend content to users and personalize nearly every aspect of their experience on the platform” (Pajkovic 2021, 3). The NRS is responsible for approximately 80 percent total hours streamed on Netflix, with the remaining 20 percent from search. The combined impact of personalization and recommendation was valued at an estimated one billion dollars per year in revenue as at 2015 (Gomez-Uribe and Hunt 2015, 5).
The NRS uses a combination of content-based filtering and collaborative filtering algorithms to recommend content. Content based filtering relies on a user’s individualized past data (their viewing history, scrolling behavior, and watch time) and makes recommendations similar to those a user has previously demonstrated interest in (Pajkovic 2021, 3). Collaborative filtering relies on larger trends among Netflix’s global users, making recommendations based on the interests and preferences of other users identified as having similar tastes (Singh 2020, 8–9). Previously, Netflix relied on collaborative filtering data from users in a specific country or region; now recommendations are gathered from algorithmically grouped “taste communities” around the world, of which there are currently over two thousand (Pajkovic 2021, 4). As Netflix seeks to commission original content around the world, its recommender algorithms have been employed at global scale, sharing data across the more than 190 countries Netflix operates in (Pajkovic 2021, 3).
The level of detail ascribed to the data Netflix collects is key to the accuracy of its recommendations. Each user’s experience of the Netflix homepage is algorithmically generated, including the rows of titles displayed, the ordering of those rows, as well as suggested titles (“Because You Watched” rows) (Gomez-Uribe and Hunt 2015, 13:04). The NRS thus consists of not one but a number of different algorithms to personalize a user’s experience of the platform. For the remainder of this essay, I focus on a specific algorithm which recommends personalized artwork to consider Netflix’s engagement with diversity through its “picturing” of diversity. What counts as “diversity” differs between countries according to the representation of social, linguistic, and cultural diversity at a local level. In this essay, diversity of content is viewed from the perspective of Netflix’s American and Australian catalogs. While studies of cultural diversity tend to focus on how accurately representations on screen reflect society (Turner 2020, 20), this needs to be combined with a consideration of the structures and people that make up a media organization, and the media’s platform specific affordances.
Personalized Artwork
In December 2017 Netflix reported the use of a new algorithm to deliver personalized artwork to its subscribers. This algorithm is capable of choosing a particular image out of several different versions to best support a given recommendation (Gomez-Uribe and Hunt 2015, 13:05). Previously, images on the Netflix site were limited to official promotional material supplied by a production company, namely movie posters or DVD cover art (Brincker 2021, 87). However, internal research showed that thumbnails of promotional posters were the major influencer on what to watch while browsing, with 82 percent of audiences deciding on the basis of these images (Nelson 2016). Nick Nelson, Head of Product Creative at Netflix, notes that users spend an average of 1.8 seconds considering thumbnail images of each title before deciding whether to view it or move on. This means there is only a very short amount of time to capture a user’s interest. Nelson (2016) observes, “we know that if you don’t capture a member’s attention within ninety seconds, that member will likely lose interest and move onto another activity. Knowing we have such a short time to capture interest, images become the most efficient and compelling way to help members discover the perfect title as quickly as possible.” 3 The personalized artwork algorithm is central to what Brincker (2021) calls Netflix’s “data-driven analytics to optimize the power of pictures” (p. 87).
The Netflix personalization algorithms are tested primarily through what are referred to as A/B tests. These tests measure the effectiveness of recommendation variants by comparing control and experimental groups of Netflix users (Gomez-Uribe and Hunt 2015, 13:09). Each group receives alternate recommendations, and user engagement and subscriber retention rates are correlated with various algorithm variants. In the case of personalized artwork, the A/B testing model tracks which images receive the most clicks in different markets (Netflix 2016). The online learning framework supporting the use of A/B testing in relation to artwork personalization involves “contextual bandits.” Netflix engineers Chandrashekar et al. (2017) explain: “Rather than waiting to collect a full batch of data, waiting to learn a model, and then waiting for an A/B test to conclude, contextual bandits rapidly figure out the optimal personalized artwork selection for a title for each member and context.” The contextual bandit model selects the artwork most likely to elicit engagement for each Netflix user based on their viewing context.
While the personalization of promotional images is not new and different film territories often have localized marketing campaigns, in the case of Netflix this material is being decided by algorithms instantly. As Lobato (2019) notes, “Netflix users do not experience the catalog as a static list or schedule, but rather as a series of interactive, personalized recommendations” (p. 251). Beyond appealing to users to simply click on an inviting image, Netflix also examines quality of engagement to avoid learning a model that recommends “clickbait” images: “ones that entice a member to start playing but ultimately result in low-quality engagement” (Chandrashekar et al. 2017). Improving engagement (the length of time viewing Netflix content) is “strongly correlated with improving retention” (Gomez-Uribe and Hunt 2015, 13:09).
Netflix’s personalized recommendations and personalized visuals means that it can respond to diverse audiences in a more agile way, but does this translate to an active engagement with diversity and inclusion? Netflix argues that its online learning algorithms or contextual bandits are “learning an unbiased model on an ongoing basis” (Chandrashekar et al. 2017). However, is this model truly “unbiased”? Hallinan and Striphas (2016) point to Netflix’s “complex alchemy of audiovisual matchmaking” (p. 117). This “complex alchemy” is generally not demystified or publicized to Netflix users. In fact, Netflix’s personalization of artwork was only brought to wider public attention in 2018 when some users began noticing the use of minor supporting Black cast in promotional thumbnails, drawing attention on social media to this perceived targeting of content based on race.
On 18 October 2018, writer and podcaster Stacia L. Brown (@slb79) tweeted: “Other Black @netflix users: does your queue do this? Generate posters with the Black cast members on them to try to compel you to watch?” Brown was commenting on a personalized promotional image of Lauren Miller Rogan’s Like Father (2018), starring Kristen Bell, Kelsey Grammer and Seth Rogan. The thumbnail on Brown’s page, however, featured black cast members Leonard Ouzts and Blaire Brooks, both of whom have only minor roles, or what Brown notes was “10 cumulative minutes of screen time [and] 20 lines between them, tops.” 4
Other users followed by posting their own examples of image personalization on social media. Tobi Aremu, a Brooklyn-based filmmaker, reported to The Guardian that the romantic comedy Set It Up (Claire Scanlon, 2018) was promoted to him with an image of supporting cast members Lucy Liu (who is Asian American), and Taye Diggs (who is African American), instead of principle cast members, Zoey Deutch and Glen Powell, both of whom are white (Sharf 2018). British romantic comedy Love Actually (Richard Curtis, 2003), with a predominantly white ensemble cast, has also been reported as featuring a promotional poster foregrounding Black actor Chiwetel Ejiofor as the romantic lead (Andrews 2018).
Perhaps the most recent high-profile example, and a case of “reverse” profiling, resulted in Nicole Byer, host of popular Netflix baking show Nailed It! speaking out. Byer was contacted by a fan who sent her a promotional poster of Nailed It! featuring only her two white, male co-stars—co-host and chef Jacques Torres and show assistant director Wes (Shepherd 2019). In a series of since-deleted Tweets, Byer wrote: If Netflix didn’t sign my checks and give me a huge platform and opportunity to showcase my comedy, I would talk about how f***** up and disrespectful this is to me a black woman. . . . I would talk about how this essentially whitewashing for more views. But they sign my checks and I’m honestly so happy and grateful to and for the show and no sarcasm I love my job and wish to keep it so I’ll be quiet (Shepherd 2019).
Byers explains her reason for deleting the tweets: “I talked to one of the execs on my show about it and the thrilling conclusion is the removal of the image and a conversation about how the thumbnails are made and selected that I’m happy with” (@nicolebyer, 29 May 2019).
Netflix has denied using race or other markers of identity as factors in its personalization data: “We don’t ask members for their race, gender, or ethnicity, so we cannot use this information to personalize their individual Netflix experience. The only information we use is a member’s viewing history” (Iqbal 2018). Rochelle King, Netflix’s vice president of product creative, adds, “In general, a person’s race, gender or ethnicity is not a great indicator of what that person will actually enjoy watching. Time after time, we have seen that great stories transcend borders and that an individual’s tastes are complex and multifaceted, going well beyond basic demographic attributes” (Nguyen). 5 And yet these “basic demographic attributes,” including race and ethnicity, are seemingly mapped onto “taste communities,” creating the potential for filter bubbles.
Alexander (2016) points to a contradiction between “the notion that we have reached an ‘on-demand utopia’ in which we are finally free to develop our own taste, and the neoliberal reality of filter bubbles” (p. 94). She argues that Netflix’s personalization system disempowers users and encourages instant gratification based on existing preferences: “we are no longer serendipitously exposed to [new and unfamiliar] films” (p. 94). Bucher (2018) affirms: “What we see is no longer what we get. What we get is what we did and that is what we see” (p. 2). Algorithmic biases are therefore evident “both in the content that is available to us and in what is not recommended” (Siles et al. 2019, 511). Described as “deceptive,” creepy’, “misrepresentative,” and creating the “clickbait” culture it seeks to avoid, personalized imagery “[plays] on the desire for Black content (recognized by the Netflix algorithm) and then utilized by Netflix (in their clickbait image production) to serve more white mainstream American culture” (Brincker 2021, 90).
Algorithmic Cultures and Bias
The transparency and accountability of algorithms that filter, hierarchize, and recommend have been scrutinized in terms of their implications and real-world consequences including discrimination and the reproduction of existing power structures (Rieder et al. 2018, 51) Theorists such as Benjamin (2019) have argued convincingly that racial and other biases are built into algorithms, perpetuating the biases of the programmers behind them: “bias enters through the backdoor of design optimization in which the humans who create the algorithms are hidden from view” (Benjamin 2019, 11). In Race After Technology, Benjamin (2019) explains what she refers to as the New Jim Code: “the employment of new technologies that reflect and reproduce existing inequities but that are promoted and perceived as more objective or progressive than the discriminatory systems of a previous era” (pp. 5–6). 6 While technology may be developed to address different forms of discrimination or bias, they may well end up reproducing them (Benjamin 2019, 47).
Noble (2018) has also highlighted how algorithms “reinforce oppressive social relationships and enact new modes of racial profiling,” which she terms “technological redlining” (p. 1). Noble writes: Part of the challenge of understanding algorithmic oppression is to understand that mathematical formulations to drive automated decisions are made by human beings. While we often think of terms such as ‘big data’ and ‘algorithms’ as being benign, neutral, or objective, they are anything but (p. 1).
Others such as Couldry and Mejias (2019) have established a connection between algorithms and “data colonialism”: a form of exploitation that “combines the predatory extractive practices of historical colonialism with the abstract quantification methods of computing” (p. 337). Algorithms construct and enforce regimes of power and knowledge, which then become further normalized. “In ranking, classifying, sorting, predicting, and processing data, algorithms . . . help to make the world appear in certain ways rather than others” (Bucher 2018, 3). Siles et al. (2019) caution even more expansively, “at stake in the establishment of algorithmic recommendation systems are the conditions for the redefinition of subjectivity and culture itself” (p. 499).
Algorithms can enact and perpetuate racial discourses, including the ways in which we think about and experience race through our engagement with various media forms (Joyrich 2009, 2). In the case of artwork personalization, the selection of minor actors of marginalized races or ethnicities being used to “sell” a program as diverse only masks the lack of diversity actually being presented, and feeds into pre-conceived ideas about what an ethnically or racially diverse audience member might want to see as part of the same “taste community.” This taste community is seemingly correlated with race by ostensibly “unbiased” algorithms.
On the other hand, some of Netflix’s more progressive and racially diverse series can end up being “whitewashed” in their promotional imagery. One of Netflix’s most popular series to date has been Bridgerton, produced by Shonda Rhimes and debuting on Netflix on 25 December 2020. The series is ground-breaking for many reasons, not least of which is its incorporation of non-white cast in lead roles in a series set in Regency era London. Bridgerton was the most watched Netflix series at the time of its debut and has been renewed for a second season in 2022. It presents an alternate history of a racially diverse society where the Queen is biracial (Queen Charlotte is played by Guyanese-British actress Golda Rosheuvel), as is the leading man, the Duke of Hastings (played by Regé-Jean Page, whose mother is Zimbabwean and whose father is English). However, in several of the promotional thumbnails for the series, not a single black actor is presented, making the series appear as another all-white Sense and Sensibility or Downton Abbey targeting viewers of traditional British period dramas. Only one thumbnail features the interracial romance at the heart of the series, between the leading couple Lady Daphne (Phoebe Dynevor) and the Duke of Hastings. A thumbnail with a much smaller image of the couple does not clearly show their faces, or allow viewers to readily ascertain their ethnicity (Figures 1–5).

Artwork Personalization of Netflix series Bridgerton (2020–2022). Period drama with an all-white cast, or intercultural romance?
Whether this promotional strategy leads to more audiences tuning in and having their expectations of a genre pleasantly disrupted, or whether it leads to disappointment and lower audience engagement, remains to be seen. In the case of Bridgerton, the show’s popularity would appear to suggest the former. Nevertheless, I argue that there is a level at which the ground-breaking and progressive nature of the series is neutralized (or worse, negated) through its promotional whitewashing, which does not encourage viewers to make a choice to watch something outside of their comfort zone. As Joyner (2016) has emphasized in relation to discovering ethnically and racially diverse content, the problem is “not in what we’re being shown, but in what we’re not being shown. . . . [I]t’s not until you express specific interest in ‘black’ content that you see how much of it Netflix has to offer. . . . [T]o the new viewer, whose preferences aren’t yet logged and tracked by Netflix’s algorithm, ‘black’ movies and shows are, for the most part, hidden from view.”
A consequence of this “picturing of diversity” is related to the fact that Netflix produces original films and television shows based on viewing data collected from audiences. Thus, recommendation systems can also influence decisions about the kinds of future programs we might see. Programs such as Orange is the New Black (created by Jenji Kohan and first airing on Netflix in 2013), and The Chair (created by Amanda Peet and Annie Julia Wyman in 2021), have been leading the way in terms of foregrounding ethnic and racial diversity in front of and behind the camera. Through their popularity, it is apparent that there is a market for Netflix to create more diverse programs for its subscribers.
Concerns over algorithmic bias, particularly in artwork personalization, intersect with Netflix’s diverse company profile in multifaceted ways. If algorithms can be programmed for taste reproduction, can they, through a diverse workforce, also be programmed for diversity?
Netflix: Sowing the Seeds
On 1 August 2017 Netflix launched the #FirstTimeISawMe campaign. Through the company’s social media channels, Netflix featured short videos with well-known creative artists including directors Spike Lee and Ava DuVernay, speaking about the importance of creative control and representation in the media for minority groups. In another video featuring its own employees, Netflix showcased the diversity of its staff, asking employees to reflect on the first time they saw themselves reflected in the media in terms of their race, sexuality or gender. Labeling themselves (in bold text on screen) as “Gay and Middle Eastern,” “Strong, Outspoken and Dominican,” and “Christian, a Mom and Asian American” (much like the descriptive tags of its micro genres) these individuals recall the powerful moments in films and television programs when they felt themselves truly represented on screen. While the #FirstTimeISawMe campaign celebrates Netflix’s own inclusive media and diverse programming, it has circulated more widely on social media to initiate a conversation about media representation, diversity, and its impact on audiences and creators.
Netflix has made a pointed decision to hire programmers and other employees from diverse backgrounds. The company has a dedicated Director of Inclusion Recruiting, Kabi Gishuru, whose job is to train employees in addition to using inclusive hiring practices to increase the company’s employee diversity, “spotting bias in the interview process, sourcing candidates in non-traditional ways, and helping hiring managers identify the perspectives missing on their teams” with the aim to create an environment, policies and practices “that not only invite people in, but when they get in they feel there is a level of investment in them” (Myers 2021). In February 2021, Netflix announced the creation of the Netflix Fund for Creative Equity, which will invest $100 million over the next five years in organizations that help underrepresented communities train and find jobs in film and television (Sarandos 2021).
In collaboration with organizations that support and promote technical, creative, and business leaders from under-represented groups, Netflix also builds diverse networks to increase its hiring pool with the aim of hiring more inclusively. It has partnerships with /dev/color, which connects Black software engineers, technologists, and executives to companies, Techqueria, which serves the largest global community of Latinx professionals in the technology industry, and TalentoTotal, a diversity and inclusion development program in the United States that promotes Afro-Latino and Indigenous (ADI) people to become business leaders across Latin America. These efforts at addressing systemic issues that have excluded particular groups from the entertainment and technology industries arguably have a flow on effect in terms of how Netflix approaches its content. Chief Content Officer for Netflix, Sarandos (2021) notes, “inclusion behind the camera exponentially increases inclusion in front of the camera, and . . . both depend on ensuring that the Netflix executives commissioning these stories are also diverse.”
In July 2020 Que Minh Luu joined Netflix as Director of Content for Australia and New Zealand. Luu is a former executive producer for the ABC where she championed diverse content such as The Heights, Diary of An Uber Driver, and Retrograde. Luu is the daughter of refugees who fled Vietnam to Australia by boat in the late 1970s. She also speaks during interviews of having a diversity and inclusion lens: “There needs to be more points of view contributing to the conversation” (Bizzaca 2021; see also Tadros 2021). Since launching locally in 2015, Netflix has made over 50 Australian titles, including co-productions with television networks. As Que Minh Luu notes, “[Netflix executives] have always been quick to make clear that being inclusive doesn’t mean that you can’t be commercial” (Bizzaca 2021). Netflix’s focus on diversity and inclusion serves a corporate agenda as much as a social one.
Beyond the corporate speak that “pair[ing] . . . culture with diversity and inclusion . . . unlocks our ability to innovate, to be creative, to solve problems,” the Netflix Inclusion report concludes with this simple statement that through diverse stories “we’re able to better entertain our current and future members” (Myers 2021). In an upbeat video published on the Netflix website, Vice President of Inclusion Strategy at Netflix, Myers (2021) comments, “My team’s vision is to equip everyone with a diversity lens, which is to say that as they do their job, they’re thinking about who’s not here. Are we gathering all the perspectives?” The idea of inclusion “taking root” at Netflix suggests that the company has invested many years in developing a strategy that is now firmly implanted as part of its culture and ethos, so much so that Netflix operates with a “diversity lens” in all aspects of its operations. The key question is how this “rooted” form of inclusion informs Netflix’s design of its algorithms and translates to audience engagement (and ultimately retention).
Netflix executives deploy all the appropriate diversity and inclusion language and is training employees in this language. The company has held workshops on topics of privilege, bias, intersectionality, and allyship, particularly in years where Black and Asian communities have been disproportionately affected by the COVID-19 pandemic and experienced hate crimes (Myers 2021). Yet, as Ahmed and Swan (2006) note, “One of the primary defences of the language of diversity is that it is more ‘inclusive’, precisely because it does not name a specific social category (such as gender, race and class). But what are the terms of this inclusion? Who is included by the term?” (p. 96).
Netflix notes that it needs to improve its recruitment of Latinx, Middle Eastern/North African, American Indian/Alaskan Native, and Native Hawaiian/Pacific Islander communities, and also its representation of the LGBTQ+ community and characters with disabilities, which currently make up only 4 percent of leads in film and 1 percent in TV series, and 1 percent of series leads, respectively (Myers 2021) Netflix has published its diversity data quarterly on its jobs site since 2013, with a plan to develop a roadmap for ongoing improvement of diversity and inclusion. Currently, women make up nearly half of its U.S. workforce (47.1%), including at the leadership level (directors and above: 47.8%). Nearly half of its U.S. workforce (46.4%) is made up of people from one or more underrepresented racial and/or ethnic backgrounds, including Black, Latinx or Hispanic, Indigenous, Middle Eastern, Asian, and Pacific Islander backgrounds (Myers 2021). As noted earlier, the Netflix Inclusion report is highly U.S. focused, with inclusion and representation outside the U.S. an area where growth is needed. The final piece of the Inclusion strategy involves the audience itself. How do audiences engage with Netflix’s personalization systems and is there room to move within these systems?
Wither the Audience?
Charges of algorithmic bias have been met by two contrasting perspectives when it comes to the agency of the audience. On one side is the argument that personalization systems result in the diminished autonomy of audiences. While ostensibly being offered a plethora of individualized options and choices, the effect of personalized recommendations is to reduce agency and autonomy by being presented with a narrower range of (similar) content (Arnold 2016, 50). If agency is defined as the capacity to act (or to have acted differently), recommendation algorithms control our ability to act by hiding certain options from view. Through algorithmic predictions, the NRS takes actions on behalf of (or away from) the user. As Arnold (2016) notes, “Although Netflix’s brand identity centers on notions of user choice, its algorithms work to actively negate choice” (p. 59).
The opacity of Netflix’s recommendation systems is harnessed to the company’s advantage. There is no real transparency or accountability about how the NRS operates or makes decisions. Furthermore, Netflix offers users “only a limited set of controls over how algorithmic decision-making shapes their platform experience” (Singh 2020, 33). Users cannot opt out of personalization features, including, and especially, receiving title suggestions. The lack of transparency makes it difficult both to analyze and to counter problematic recommendations derived from these systems (Singh 2020, 6).
The other side of the debate suggests that there is a correlation between the use of SVOD platforms and an increase in the quantity and diversity of content consumed, despite recommendation systems offering a narrower range of content over time. As Limov (2020) suggests, “users are transformed by recommendations, even as their participation transforms the algorithms in turn” (p. 6307). Limov (2020) observes, “recommendation systems can be conceptually understood as channelling users” attention to the (somewhat) unfamiliar, adding a dimension of discoverability that improves the accessibility of content on streaming platforms’ (p. 6307). Gillespie (2014) adds, “recommendation algorithms map our preferences against others, suggesting new or forgotten bits of culture for us to encounter” (p. 167).
These viewpoints are not polar opposites, and research in this area has been limited by a lack of knowledge on how users actually respond to recommendation algorithms on a daily basis or incorporate them into their lives (Siles et al. 2019, 499). In the Australian context, cultural theorist Turner (2019) signals more broadly that there is a “gap in our knowledge of how individuals and households consume television, across platforms and devices, in domestic spaces” (p. 222), and this requires us to adapt modes of audience research employed in earlier studies of television audiences. It is not yet clear what kind of impact recommender algorithms can or will have in terms of transforming film and television cultures and developing greater engagement with more diverse content. A more balanced approach between the two opposing views suggests that the relationship between users and algorithms can be framed in terms of “mutual domestication”: While algorithms participate in the maintenance and perpetuation of certain cultural codes, they also learn from them and can be incorporated into our daily lives . . . differently, shaping culture differently. The relationship between platform, technology (algorithm) and people is cyclical. We watch on the basis of recommendations, but then these perpetuate certain other recommendations’ (Siles et al. 2019, 516).
Within these cyclical processes there are a range of actors, including the programmers and engineers who develop these systems, the users who interact with them, and the company that deploys them (Pajkovic 2021, 3).
While resistance against the technical aspects of recommendation systems is not easily achievable, it may be possible to resist the cultural biases inscribed in Netflix’s recommendations, “expressed in the constant recommendation of content that users consider stereotypical” (Siles et al. 2019, 511). Audiences can also actively exercise agency through their social networks. Recommendations from friends, peers, and other social networks also shape this relationship (Frey 2019, 167), as the example of Bridgerton shows, despite the persistent whitewashing of its promotional imagery. Algorithms thus form part of broader situated practices of sociality. Bucher (2018) posits the notion of “programmed sociality” as a heuristic. Within certain circumscribed parameters, algorithms are dynamic rather than fixed and participate in wider networks of sociality, both human and non-human (Bucher 2018, 4).
Thus, while it may not be possible to actively circumvent the Netflix recommender algorithms to see more diverse content on a user’s home page, it might be possible to resist aspects of personalization through other forms of “programmed sociality.” The way Netflix users relate to the media they are presented with is not only technologically determined. While algorithms may express the practices and biases of their designers, they are also shaped by users’ social and cultural codes, and in the case of Netflix, by its diverse company profile, at least in the United States.
Conclusion
In this essay I analyzed the relationship between streaming platforms, recommender algorithms, and cultural diversity to consider the question of whether streaming services present more diverse stories and facilitate a greater engagement with diversity and inclusion than traditional “free-to-air” television, or whether they are simply better at marketing and promoting diversity to audiences. I focused specifically on the Netflix Recommendation System, and in particular Netflix’s personalization of artwork.
Where free-to-air television has been slow to cater to diverse audiences (with notable exceptions, for example, in the Australian context, the nation’s public broadcasters, the Special Broadcasting Service (SBS) and the Australian Broadcasting Corporation (ABC)), other platforms, including streaming services, have to some degree filled a gap. Competition between a growing stable of streaming services arguably leads to more diversity, as platforms themselves have been required to engage with different audiences to grow their subscription base. Streaming platforms do not have the limitations of scheduling that free-to-air television has; however, they must capture audiences’ attention as soon as they connect to the platform (Ranaivoson 2019, 14). Recommender systems are designed to assist users in making choices based on their previous selections. This does lead to biases in what is presented, and indeed, in the case of artwork personalization, how they are presented.
I have presented the argument that recommender algorithms take actions on behalf of (or away from) users, who are not necessarily given access to the full range of Netflix’s catalog. While there is greater diversity being presented on the surface, the question remains whether this results in sustained engagement with what is being watched, with users transformed by the algorithmic recommendations to extend their viewing habits, or whether recommender systems merely reinforce existing consumer tastes, perpetuating consumption of similar products with only the “illusion of diversity” (giving the impression that the Netflix catalog is much bigger than it actually is) (Ranaivoson 2019, 112). Further audience research needs to be conducted in this area.
Beyond the level of on-screen representation, the value of diversity and inclusion can be found in having a greater pool of creative talent to draw from, to create stories that will have relevance to a broader cross section of the population, and to tell more innovative stories. As the Screen Australia (2016) report Seeing Ourselves: Reflections on Diversity in TV drama notes, the lack of diversity and inclusion: is limiting the relevance of our industry and our most popular forms of cultural expression. It is having commercial implications, as audiences seek relevant content elsewhere in material produced overseas. And it is undermining our ability to innovate and connect with the storytelling potential of our increasingly diverse population.
Netflix has acknowledged the value of a diverse workforce by placing its Inclusion strategy at the heart of its operations. From a cultural perspective, we can also point to a broader shift in audience’s media consumption toward more diverse content; that is, reflecting the progressiveness of contemporary audiences and their commitment to diversity and to a more accurate representation of the make-up of their society.
The power of algorithms, as Bucher (2018) notes, is “through the kinds of encounters and orientations algorithmic systems seem to be generative of” (p. 3). Algorithms are relational, processual, and cultural (Striphas 2015). The picturing of diversity may for now be algorithmically generated but more diverse content is slowly making an appearance if you know where to find it (or it will eventually find you).
Footnotes
Declaration of Conflicting Interests
The author declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author received no financial support for the research, authorship, and/or publication of this article.
