Are Recommendation Algorithms Killing Content Diversity? A Netflix Case Study.

Notorious for its hyper-personalized recommendation algorithm, Streaming Video On Demand (SVOD) platform Netflix uses Artificial Intelligence (AI) to curate a an individualized experience for users based on user data. While individual recommendations may make 'discovering' programs that align with user preference easier for some, others are left to deal with issues such as filter bubbles and content misrepresentation.

Join author Erica Turner as she investigates the infamous Netflix Recommendation System, determining how content diversity might be impacted by artificial intelligence softwares.

One could compare the introduction of Artificial Intelligence (AI) technology into society to the 1958 film The Blob. Just like the extraterrestrial carnivore from the movie, AI technology has been silently evolving itself, becoming a sizeable threat to the livelihoods of people in both our local and global communities.

In the film, two teenagers discover that some sort of gelatinous-looking creature is responsible for the death of an elderly man in town. The teenagers try to warn officials about the apparent killer alien on the loose, while the blob grows larger and more powerful with each victim it consumes. As much as the kids beg and plead for police to take them seriously, they fail to make a convincing argument and are ultimately dismissed. The blob becomes so powerful that it attacks hundreds of people in the town’s movie theater, leaving officials staring on in awe as people run for their lives.

While the blob is eventually tamed, there is a precautionary tale to be taken out of this sci-fi film as it relates to AI technology. Like the blob, AI isn’t very well understood by the general public (or even the engineers who have created it themselves), it has the capacity to evolve itself through self-learning programs, and it’s impact in society has gone majorly undetected by our government officials.

Unlike the film, however, there are of hundreds of researchers hoping to bring the current and potential issues with AI to light. I aim to join in the effort to unmask AI’s facade of perfection by investigating it’s affordances and pitfalls as it relates to the entertainment industry.

The threats that AI create for the entertainment industry have been very prevalent in American news cycles over the past year, since one of the major demands being fought for in the 2023 SAG-AFTRA strike was the regulation of AI usage in writers rooms and the regulation of deepfake technology as a replacement for on-screen performances. Without regulation, these systems can replace the jobs of screenwriters and actors alike, and may also draw on prior works without properly compensating the original writers or actors. The algorithmic construction of projects also means the creation of entertainment that reflects cultures in a computerized way, thus contributing to culture from artificially constructed projects that lack the human touch that typically ascribes meaning to works of cinema.

Also in a ‘Blob’-like fashion, the way that artificial intelligence is used in the entertainment industry to promote projects, rather than aid in their creation, is a utilization of AI that has been widely disregarded since it’s effects aren’t necessarily ‘in-your-face’- i.e. you might be able to spot a deep fake, or you might lose your job as a screenwriter, but you won’t notice the immediate effects of a computer recommending something to you. 

The use of recommendation algorithms to popularize certain projects among online communities might feel normal to the average Gen-Z or Gen-Alpha user, but home entertainment recommendations were not always created by a computer. To understand how media companies might be taking advantage of these recommendation technologies, I performed a case study on the Netflix Recommendation System (NRS) to assess whether or not it was capable of promoting diverse content. To best contextualize how the NRS came to-be, let’s take a look at the world before Netflix, and a world pre-recommendation algorithm. 

TV Guide- America’s Long Vigil, Jan 25-31 1964, TV Guide, Triangle Publications, Philadelphia. Pa, USA, 1964. Pp A50-A51. https://flic.kr/p/i1bMiT 

Before recommendation algorithms, there was the TV-Guide.  The TV-Guide was a weekly magazine that informed audiences of what content would be broadcast on what channel at what time; featuring the title and synopsis of the show. The TV-Guide allowed viewers to peruse the entire catalog of content and select when to tune in and catch the next episode of their favorite show, or perhaps view a rerun of a beloved nostalgic movie.

Programs could purchase advertisement slots or be featured in stories as a part of the traditional ‘magazine’ portion of the guide. The TV-Guide was a tangible thing that you could flip through and observe. Recommendations through the guide were made in two ways; the first being through descriptions that viewers could read and determine if the program fits their entertainment desires, the second being through advertisements that highlight certain programs, such as the promotion of “Ralph’s Story” in the photo above. The TV-Guide magazine was discontinued in 2008, however it can now be found online at the TV-Guide website. The TV Guide is flirting with its inevitable demise, as television viewers switch from broadcast to video entertainment platforms such as streaming services and YouTube.

As technology developed, the VHS player became a game-changer for home entertainment media. No longer being tied to broadcast’s strict air times, audiences could select a VHS (or eventually a DVD) of their preferred show, and could watch it at any time. Founded in 1985, Blockbuster became the most popular at-home entertainment rental store in the United States. In Blockbuster, titles would be organized by genre, and would be presented face-forward so customers could see the cover image for each show’s case. Each copy of a single title had the same exact case, with the same cover art/photos and information made available to describe the content on the tape. Some tapes, however, would have a generic Blockbuster cover with a sticker of the title, typically used when the original cases were damaged. 

AP Photos/ Kevork Djansezian, https://www.businessinsider.com/best-things-blockbuster-2018-8 

Another component of being in a physical Blockbuster store is that recommendations could be given by ‘product experts’, or people who understand the catalog in the store and can direct you to a relevant title based on your preferences.

The benefit of speaking with a product expert is that a certain level of trust is garnered from their supposed expertise and the human interaction. Assistant professor at the UT Austin School of Information, Min Kyung Lee, has extensively conducted research on human-computer interaction to understand how AI can strengthen human decision making. In a 2018 study, her research suggested that people found human decisions to be perceived as more trustworthy than algorithmic decisions, due to the algorithm's lack of subjective judgment and the ‘dehumanizing’ process of being assessed by a machine.  So that raises the following  question:

If humans generally perceive algorithms to be less trustworthy, how does Netflix save $1 billion dollars a year using AI to ‘judge’ users on user data that powers their own experience?

This question can be answered by understanding why the shift from the tangible recommendation systems such as the TV-Guide and the home entertainment video store environment, to digital recommendations like the NRS occurred in the first place. Digital recommendations reflect a desire for convenience from consumers, and a corporate desire to target consumers on a more personal level. 

Regardless of how humans feel about AI decisions, the digitization of home entertainment video libraries is able to increase the audience’s autonomy, as they can now watch whatever title they desire, at whichever time they prefer, from anywhere they have an internet-connected device. Businesses use these data-driven algorithms to match users with content that fits the tastes of the user, personalizing their experience on levels that might even be shockingly accurate. 

But just how personal is too personal? 

Netflix is known as a leader in recommendation algorithm softwares, providing immense personalization for users all over the globe. Let’s review a brief history of the Netflix Recommendation System (NRS) to understand how it came to be, and how titles are chosen for each subscriber.

As a subscriber-based business model, user experience is integral to retaining subscribers and keeping Netflix profitable. The Netflix Recommendation System (NRS) began during a 2006 contest held by Netflix, titled “The Netflix Prize”. If anyone could predict user ratings 10% better than Netflix’s previous algorithm ‘Cinematch’ could, they would win a one million dollar grand prize. The contest was won in 2009 by a team of AT&T researchers, 3 years after the contest began. Plenty of addendums and changes have been made since 2009, and the combination of these changes and algorithms make up what we know to be the Netflix Recommendation System today. 

So how does it work?

Well, Netflix keeps information about their algorithms pretty secretive, so I had to construct an understanding of the system based on the information that is known, with an understanding that other unknown factors are likely at play. Most recommendation algorithms use content-based filtering (showing items based on your data/algorithmically determined interests), and collaborative-based filtering (showing items based on other profiles that are algorithmically determined to be similar to yours). Netflix uses these systems in a combination of other processes to select which titles to promote to specific users. 

A useful tool in outlining the NRS is the presentation “Personalization at Netflix: Making Stories Travel”, which was presented by ex-lead researcher of machine learning at Netflix, Sudeep Das. He lists the four ‘main ingredients’ that make up a user's personalized experience as their viewing history, thumbs ups/downs and + My List adds, title details, and members with similar tastes. 

Upon the creation of a new Netflix account, users are asked to select three TV shows or movies that they like, and from there the system uses this initial data to make guesses based on popularity. As users watch more on the platform, the algorithm will identify them with ‘interest groups’- groups of users all around the world that are watching similar content, which is commonly referred to as ‘taste clusters’ or ‘taste communities’. 

 Let’s look at the Netflix home page.

Netflix Home Page, Niko Pajkovic, 2022, https://journals.sagepub.com/doi/full/10.1177/13548565211014464. 

In this sample home page, we can visualize Netflix’ recommendations for this particular user. The first step in creating the Netflix homepage is to decide which titles the member should see, based on data attributed to their profile. Then, the row headers that would be most appealing to the member are selected to represent each row. After this is completed, the order in which the titles are presented in the row is chosen (i.e, selecting if The Age of Adaline should go before The Best of Me in the ‘emotional dramas’ row). The rows are then placed into the most prominent spots based on their relevance (i.e. putting the ‘emotional dramas’ row above or below the ‘movies for hopeless romantics’ row for relevance reasons). Lastly, duplicates are removed from each row. 

A few addendums to the NRS since this presentation are the customization of thumbnails and the Global Top 10 feature. 

Varying Netflix Thumbnails, Niko Pajkovic, 2022, https://journals.sagepub.com/doi/full/10.1177/13548565211014464.

Thumbnail variants for each title are assigned to certain ‘taste profiles’. Published AI researcher Niko Pajkovic conducted a study to analyze different algorithmically constructed homepages based on stereotypical interactions from three different kinds of viewers. He took on the persona of a ‘Die-Hard Sports Fan, a ‘Hopeless Romantic’ , and a ‘Culture Snob’. The difference in thumbnails for each persona is shown above. These thumbnails vary based on what the algorithm believes will be the most visually enticing for each type of user.

The ‘Global Top 10’ feature takes total viewer hours from all countries where the programming is available to determine a weekly ranking of shows and movies on the platform. Theoretically, the ‘Global Top 10’ is the only category that would be completely the same for users all over the world. This feature was created to help users explore content outside of their central interests. 

The way Netflix recommends programs is very different from the TV Guide or Blockbuster. One of the major differences is that the entire library of content is largely hidden from view on Netflix, as titles are only available when searched for, by genre, or when the algorithm pushes a title onto the user’s Homepage. Another notable difference is that the presentation of titles varies based on the user, whereas the TV-Guide and Blockbuster had a singular synopsis and a singular version of cover art to promote to every user. Personalization may make it easier for viewers to find relevant shows, however, the ability to target users so accurately may not be such a great thing. 

A problem with hyper-individualized recommendation algorithms is that they can create what are known as filter bubbles. Filter bubbles refer to the phenomenon of user’s being isolated from content that they haven’t yet expressed interest in. 

For example, If a person watches a lot of movies about sports, filtering algorithms will learn that the user enjoys sports content, and will continue promoting sports related programming. However, this person may also have a love for baking shows. Because the algorithm never provided baking-related content for the user to see in the first place, the algorithm has no data indicating an interest in baking, and the user will continue seeing mostly sports-related content options. 

These hyper-personalized algorithms threaten the creation of echo chambers, which is the isolation of recommendations that do not expose users to views or interests other than what the algorithm has deemed to be their specific interests.

Another issue with personalization algorithms like the thumbnail algorithm, is that it can present a facade of content diversity. For Netflix specifically, certain titles can be misrepresented in an effort to get users to stop and click on the program. For example, if you look back at figure 4, you’ll see that the Outer Banks thumbnail for the ‘Die-Gard Sports Fan’ features two teenagers with surfboards. This thumbnail would indicate to users that the show has something to do with surfing, but in reality, surfing has nothing to do with the actual plot of the show. These deceptive practices are not only frustrating for viewers, but leads subscribers to believe that Netflix holds a vast library of unique and interesting content when in reality, these titles are being misrepresented in order to uphold a facade of diverse content. 

Thumbnail artwork algorithms are also proven to have targeted users based on race in misrepresentative ways. In 2018, many Black Netflix subscribers noticed their thumbnails showing black characters from shows where these characters were minor/ rather irrelevant.

Brown, Stacia. Twitter Post. October 17, 2018. https://twitter.com/slb79/status/1052776984231718912 

This tweet is one example of many that were taken to Twitter by frustrated subscribers. All programs are competing for your attention, as it takes about 1.8 seconds for a person to decide if they are going to stop on a title. When a thumbnail participates in this sort of ‘racial click-bait’, users are left feeling tricked.  

In Blockbuster, titles had the same cover photos (unless the cover was destroyed and had to be replaced with a generic Blockbuster one), therefore they were almost guaranteed to accurately represent what would appear in the show. There was no incentive to change how the show is portrayed, as the way to target users was to advertise to a large general audience, not to take the show and misconstrue it in a way that appeals to specific people. The NRS does not collect data surrounding race, yet its algorithm has produced a superficial visual diversity that needs to be addressed. 

These misrepresentations also are not universal; in the Like Father case, users whose algorithms construct thumbnails featuring Kristen Bell and Kelsey Grammer are not going to share the frustrations of the woman whose algorithm made it seem like the black characters were more significant to the story than they actually are.

Assistant Professor of Communication Studies at Baruch College, Stuart Davis, outlines an additional negative factor surrounding recommendation algorithms; the concept of ‘corrupt personalization’, where companies utilize their algorithms to promote their own incentives over the needs and desires of their consumer base. I found an alarming trend in the Global Top 10 feature, where Netflix may be promoting its Netflix Originals over the legitimate ‘Top 10’ shows. 

Turner, 2024.

While the TV-Guide and Blockbuster might have had paid featured sections highlighting specific titles, these promotions did not mean that the other titles would lose their chance to be presented to a potential viewer (i.e. it would not ‘boot’ a title off of the visible content library). On Netflix, however, not all titles are shown. In this sense, all titles on the homepage are fighting to be promoted. When Netflix claims that their Global Top Ten includes a decent amount of their original content, this data needs to be supported, or else it is evident that they are abusing this feature for their own benefit. 

While trying to support this finding, I came across two issues- the first being that ratings are off viewing hours- what if users who enjoy the show are watching multiple times? And what does that mean for someone who watched 40 mins of a movie, but couldn’t get through it because they thought it was terrible? 

While conducting popularity based on hours viewed has issues in its own regard, the major problem I found comes with the fact that Netflix doesn’t use an outside party to collect and report their data. 

In the analog times, TV Guide and Blockbuster would have used Nielsen ratings to deduce popularity, thus recommending or diverting viewers from watching certain shows/films with legitimate consumer reported data. Netflix using their own data to promote a ‘top ten’ of what they have to offer in their library, then putting their original content in the majority of the top 3 weekly rankings, is problematic. If their original content truly was in the top three that often, there would be no issue. The fact that they are reporting their own data, however, means that a conflict of interest could very well be present, and I would argue that it is present, based on the #1 spot being a Netflix Original 87% of the time over the past 15 or so weeks.

So, what's the answer;  Are recommendation algorithms good or bad for content diversity?

The answer is that it could be good for content diversity, but under a few conditions. 

In the case of Netflix, the library holds content with diverse stories; there are roughly 2,000-5,800 titles available depending on country, with at least 76,000 micro-genres in the United States that prove to producers there is an audience for very niche genres (ex: “TV Dramedies Featuring a Strong Female Lead”, or “Understated Romantic Independent Dramas”). Netflix is also known to produce content that is ‘45 degrees shifted’ from the typical programs you would find in the macro-genres, meaning they want their content to have a slight edge to set them apart from their competitor. 

Maintaining a library with diverse stories, however, does not equate to these stories being promoted to users. 

The existence of micro-genres has meant that there is evidence of very specific groups that Netflix would want to target, requiring an intense variety of stories. However, users may unequally receive accurate recommendations based on who they are and how the algorithm interprets their data. Algorithms that train towards similar content may consequently create filter bubbles, misrepresent content, and ultimately become a vehicle for corrupt personalization to ensue. An algorithm that is so individualized it allows the interests of corporate greed to bleed into your unique profile is arguably where the line is drawn in determining how much personalization is too much personalization.

The conditions that would make recommendation algorithms good for content diversification is if they are subjected to regulations that would ensure they produce representational renderings of their content, and if they outsource their data collection to ensure that their claims of popularity are legitimate, not self-serving.

The Yesterday lawsuit has set a precedent that shows us how misrepresentation of media projects has previously been defended in court. The film Yesterday features actress Ana de Armas in the trailer, yet she is nowhere to be found in the film. Two De Armas superfans who were duped when their favorite actress did not appear in the film took their frustrations to court, claiming that the studio was participating in false advertisement. Universal Pictures’ defense in the case was that trailers are ‘works of art’ and therefore are subject to free speech under the First Amendment. The court ruled in favor of the superfans, citing that trailers are ‘commercial speech’, meaning they must abide by false advertising law. 

If Netflix finds themselves in a similar case, any claims made that their advertisements are works of art rather than false advertising will likely be ruled against in a court of law. The main difference between Universal Studios’ Yesterday lawsuit and this hypothetical Netflix lawsuit is that Netflix could potentially deflect blame for inaccurate trailers or thumbnails onto their recommendation technology. Enforcing regulations that hold companies liable for false advertising, even if artificial intelligence plays a role in ad creation and distribution, will force corporations to ensure that their systems do not abuse user data when deciding how to suggest content to viewers.

Utilizing this Netflix recommendation algorithm case study can also help us to understand implications on a larger scale of unregulated technology. Section 230 of the Communications Decency Act removes liability from internet platforms, making it so that they are not treated as the publisher or speaker of information found on their site. Section 230 also allows for sites to moderate content as long as the service is removing/altering content ‘in good faith’. 

Eric Goldman, a professor at the Santa Clara University of Law, claims that Netflix should be protected under Section 230 when it comes to third-party content on the site, even if Netflix has purchased the copyright of these works.  Professor Goldman was referencing Section 230 when discussing Netflix’s defamation lawsuits, however, should Netflix accept this protection under Section 230, it would then also protect its recommendation algorithms from any problems that could arise. 

       Problems that have arisen from recommendation algorithms on other platforms, unfortunately, have had lethal consequences.

In 2022, 14 year old Molly Russell committed suicide, and investigators found that the Pinterest algorithm had consistently been suggesting her with topics such as ‘10 depression pins you might like’, or ‘suicidal quotes’. 

Another case against recommendation algorithms comes from the family of Nohemi Gonzalez, a 23-year-old woman from California shot dead during a 2015 rampage by Islamist militants in Paris. The family claimed that YouTube, through its computer algorithms, unlawfully recommended ISIS recruitment videos to certain users. 

While Netflix’s artificial intelligence has not been tied to any deaths as a result of the platform’s algorithm, it would not be shocking if cases similar to Russel or Gonzalez someday arose from Netflix’s personalization technology.

Especially from a global perspective, content diversity isn't the only issue when it comes to recommendation algorithms. Outlined by international media critic Thomas L. McPhail through his description of ‘electronic colonialism theory’ (also known as ‘digital colonialism’), countries that rely on America, or any dominant country, for their technological hardwares and softwares are subject to “a set of foreign norms, values, and expectations that… alter domestic cultures, languages, habits, values, and the socialization process itself”. We can already see evidence of the algorithm’s presence in foreign countries shaping local cultures, as Netflix’s dominance in Latin America has created a larger desire for content in the drama and comedy genres, disrupting the local genre preferences.

In a similar fashion to Netflix promoting its original content through algorithms that supposedly are unbiased, other tech companies could just as easily obscure their data and promote content globally under American-based incentives. Taking away the sovereignty and autonomy to produce global cultures of our own, and not a computerized rendering nor construction of our cultures takes away from the human emotion and subjectivity that goes into creating culture and community. 

As Author Aldous Huxley predicted in his novel Brave New World, people will ‘come to adore the technologies that undo their capacities to think’, as succinctly summarized by author Neil Postman in his novel, Amusing Ourselves to Death. If the global community is ignorant of the ways in which AI recommendation algorithms interact with our communities, even on seemingly insignificant levels such as the protection of diverse content promotion on streaming services, we risk distracting ourselves away from human-built cultures and marching straight into a future where computers determine our values and interests. 

All in all, recommendation algorithms have the potential to be good for content diversity, but to live in a world where they are good for content diversity, it is important we continue improving their relationship to our communities by paying attention to their impact and pushing to regulate the ways in which they are utilized.

It also might be a good idea to maintain separate profiles from your family members on Netflix if you don’t want your dad’s hyperfixation on war documentaries or your sister’s reality TV kick to clog up the recommendations on your Netflix homepage!