If...Then Algorithmic Power and Politics

However, the hacker identity was not always a part of Facebook’s culture. The external hacker brand first had to be created. In 2008, cultural and employment branding manager Molly Graham was hired to create and tell the company’s story. Graham describes how the defining moment for Facebook’s identity came with a blog post by Paul Buchheit, who joined Facebook from FriendFeed in 2009. The post reframed hacking as “applied philosophy,” a way of thinking and acting that “is much bigger and more important than clever bits of code in a computer” (Buchheit, 2009). According to Graham, “When Paul wrote that post, he wrote the story of Facebook’s DNA” (Firstround, 2014).

Facebook’s DNA is not just externally communicated but internally communicated and “installed” into every new Facebook employee, as one of Facebook’s engineers put it in a blog post (Hamilton, 2010). Newly hired Facebook engineers have to go through a six-week boot camp in which they learn almost everything about the software stack and are exposed to the breadth of code base and the core tools for engineering. Moreover, as Andrew Bosworth—the inventor of the Facebook boot camp and a long-time Facebook engineer—puts it, the boot camp is a training ground for the camper to be “indoctrinated culturally.”4 New hires are told from the very beginning to “be bold and fearless,” an ethos that is at the heart of Facebook engineering.5 Just as the company and code base are always changing, so are their engineers expected to change. Whatever they do, Facebook engineers are continuously told to move fast. In terms of the boot camp, moving fast means starting to work on actual code right away, and there are expectations of publishing code to the live site within their first week. Moving fast is also an important part of Facebook hackathons, another cornerstone of Facebook’s hacker culture. Hackathons are organized at Facebook every six weeks for engineers to conceive of new products in a short and concentrated timeframe. Some of Facebook’s most popular products emerged from a hackathon, including the “like” buton, Timeline, and Chat.

70

I F . . . T H E N

Ultimately, culture is embodied. As understood by Raymond Williams (1985), “culture” refers to the meanings and values implicit and explicit in particular ways of life. As a news report on the Facebook boot camp suggests, Facebook counts on its engineers to be as “nimble as ninjas, always ready to change course” (Bloomberg, n.a.). According to Pedram Keyani, engineering director and company culture advocate at Facebook, culture needs to be nourished to evolve, and it needs to be embodied by the management. Being bold, fast, and open, focusing on impact, and building trust constitute the core company values and tenets of the culture at Facebook, allegedly perfectly embodied by its founder Mark Zuckerberg (Keyani, 2014). On the company’s 10th anniversary in early 2014, Zuckerberg reflected back on what he called “an amazing story so far,” pondering why, of all people, they were the ones to build Facebook. After all, he says, “we were just students.” The quick answer: “we just cared more about connecting the world than anyone else. And we still do today” (Zuckerberg, 2014). The ethos and rhetoric of openness is ubiquitous at Facebook, despite or, perhaps precisely because of, continuous allegations to the contrary.6 Zuckerberg is celebrated for openly answering any questions employees may have at a weekly company Q&A; products are launched “openly” all the time without “secrecy”; and, when Facebook moved from its headquarters, most of the walls inside the building were pulled down to support and reflect Facebook’s “open communication environment” (Keyani, 2014).

Mark Zuckerberg’s T-shirt, the moving furniture, and the open ofice space are not simply allegorical accessories. They are not meant to provide a stroke of cultural paint over an otherwise technical canvas. Mark Zuckerberg’s contempt for mundane decisions should not simply be dismissed as empty PR talk; nor should the refurbishing of Facebook’s headquarters be taken as a token of style. Then, why talk about a T-shirt and hacker culture? From the perspective that culture is a whole way of life, technologies are integral to culture, not separate from it (Slack and Wise, 2002). Culture maters to an understanding of software and algorithms inasmuch as they are inherently intertwined and co-constituted (Pickering, 1995; Slack and Wise, 2002; Williams, 2005). Indeed, as computer historian Nathan Ensmenger suggests, “software is where the technology of computing intersects with social relationships, organizational politics, and personal agendas” (Ensmenger, 2012; Haigh,  2001). Nowhere does Mark Zuckerberg’s personal agenda of making the world more open and connected, Facebook’s organizational goals of maintaining engaged users to ensure revenue, and users’ desire to see stories that interest them and stay connected to the people they care about intersect more with the power of  software to frame the world as it does on Facebook’s news feed. As a starting point, therefore, we need to acknowledge that what Facebook calls the hacker culture is deeply engrained in the code and products that emerge from the cultural life of its  engineers. While Mark Zuckerberg claims the T-shirt keeps his goals and motivations intact by eliminating one more decision from his life, it also says something about the ways in which the CEO of the world’s largest social networking site thinks about human autonomy and decision-making. Similarly, perpetuating a work environment in which moving furniture is part of the everyday life of Facebook engineers or tearing down walls for meetings to happen anywhere at any time is much more that a design decision made by an interior architect. They are the physical manifestations of the “under construction” ethos that permeates the culture of engineering at Facebook. As John Tenanes, who oversees Facebook’s buildings as its director of real estate, tellingly says about the physical layout Facebook’s headquarters: “It’s designed to change thinking” (Hardy, 2014).

Life at the Top

71

The engineering of the engineer at which Tenanes hints clearly challenges the idea that artefacts are socially constructed, as if the social was already a given. Engineers do not simply engineer artefacts as if their intentions and ideas were the results of discursive and deliberate processes. These deliberations are always already imbued in materiality, whether it is the architecture of the headquarters, a piece of clothing, or the infrastructure of the work place. Just as the cultural dimension of computing maters, so does the material dimension of the culture of computing. It becomes apparent that thinking about the power of software and algorithms in terms of cause and determination will not do. While STS scholars have long argued for the necessary entanglement of materiality, practice, and poli tics (Gillespie, Boczkowski, & Foot, 2014: 4), there is a tendency still among communication scholars to frame technology as an outcome of culture. As Leah Lievrouw suggests, communication scholars only secondarily, if at all, consider material artefacts and devices to have anything resembling power (2014: 24). Indebted to the STS ideal of co-determination and the mutual shaping of technology and cultural practices, while at the same time explicitly opposing the preference for the social/cultural side of this sociotechnical duality, in this chapter I seek to atend more closely to the physical design and configuration of the (im)material architecture of the news feed. Echoing the idea of co-production or the dynamic relationship between the material and social, power is no longer seen as an abstract “force” or institutional “structure” but, rather, as instantiated and observable in the physical forms of social practices, relations, and material objects and artefacts (Lievrouw, 2014: 31).

Toward a Diagrammatics of News Feed Algorithms

My interest in questions of power stems from the view that power is increasingly materialized in and through software, as software and code have become ubiquitous, slipping into the “technological unconscious” of everyday life (Lash,  2007; Thrift, 2005). I hold that algorithms have productive capacities not merely by mediating the world but through their delegated capacities to do work in the world and by making a diference in how social formations and relations are formed and informed. Seting out to explore power through software and algorithms, indeed, seems to be a daunting and, perhaps, even impossible task. However, it is not my intention to investigate power as a totalizing force that can be located and described once and for all. As discussed in previous chapters, algorithms are never powerful in one way only but power always signifies the capacity to produce new realities. What we need, therefore, are diagrams or maps of power, tracing how and when power operates and functions in specific setings and through specific means.7 Following Deleuze in his reading of Foucault and the assertion that “every society has its diagram(s)” (2006: 31), what is at stake here is thus a diagrammatics of algorithms, understood as the cartography of strategies of power. Here, Foucault’s analysis of the architectural, technological, and conceptual nature of diagrams provides a useful methodological and analytical framework for examining Facebook’s news feed.

72

I F . . . T H E N

Foucault’s architectural model of power usefully highlights the ways in which spaces are “designed to make things seeable, and seeable in a specific” way (Rajchman, 1988). It ofers a useful avenue to analyze algorithms in terms of architectural structuring in which embedded technically means having the power “to incite, to induce,  to seduce, to make easy or dificult, to enlarge or limit, to make more or less  probable” (Foucault quoted in Deleuze,  2006:59). For a mapping of power, Foucault’s notion of government and governmentality also ofers elaborate concepts for the kinds of architectural shaping that he first described in Discipline and Punish (1977), as it points to the multiple processes, measurements, calculations, and techniques at play in organizing and arranging sociality. Analytically, this implies a focus on the ways in which algorithms arrange and organize things, its specific techniques and procedures, and the mechanisms used in processes of individual and collective individuation.8

I argue in this chapter for the importance of revisiting the idea of the technical and architectural organization of power as proposed in the writings of Foucault, by highlighting an analytics of visibility. Becoming visible, or being granted visibility, is a highly contested game of power in which the media play a crucial role. While Foucault did not connect his theory of visibility specifically to the media, the framework he developed in Discipline and Punish helps illuminate the ways in which the media participate in configuring the visible as oscillating between what can and should be seen and what should not and cannot be seen, between who can and cannot see whom. Examining new modalities of visibility, thus, becomes a question of how and when something is made visible rather than what is made visible, through which specific politics of arrangement, architecture, and designs. Although Foucault’s writings focused more on the “how of power,” I want to suggest that the how is inextricably linked to the when of algorithms as discussed in the previous chapter. If the how of power are “those practices, techniques, and procedures that give it efect” (Townley, 1993: 520), then the when, in this case, refers to the aspects that are rendered important or unimportant to the efectuating of these procedures at diferent times. In this chapter, I investigate the notion of mediated and constructed visibility through a close reading of the news feed and its underlying algorithmic operational logic. I argue that Foucault’s idea of an architecturally-constructed regime of visibility as exemplified in the figure of the Panopticon makes for a useful analytical and conceptual framework for understanding the ways in which the sensible is governed in social networking sites.9 The intention is not so much to ofer a definite account of the role played by Facebook in capturing the world in code but to open avenues for reflection on the new conditions through which visibility is constructed by algorithms online.

Life at the Top

73

The News Feed

Launched in 2006, the news feed is one of Facebook’s primary and most successful features. The news feed, placed right in the middle of a users’ home page, is a continuous stream of updates, serving users stories of what their friends have been up to. Since its conception, the news feed has changed its design and functionalities so many times that they launched a series of blog posts in August 2013 called News Feed FYI so that people could more easily keep track of the most important changes.10 In its twelve years of existence, the news feed has gone from being a single feed of reverse-chronologically-ordered updates to becoming fully edited and governed by machine learning algorithms. In an older version anno 2009, the news feed was organized as two distinct feeds, separated by easily recognizable tabs, whereby users could choose between viewing the “live feed” or the “top news.” When the news feed revamped into a single algorithmically governed feed in the autumn of 2011, Facebook started showing real-time updates in a feed called Ticker, placed on the right-hand column of the Facebook homepage. While nothing stays the same on Facebook, the basic layout of a single filtered news feed at the centre and a ticker on the right has more or less persisted since.

Since much of the empirical material used in this chapter is based on data collected in the time period of April–August 2011, using my own personal Facebook profile, it is important to be clear about the state of the feed and the design at that particular point in time. In this time period, the news feed was divided into two diferent versions, the default “top news” and the “most recent” feed. According to the Facebook Help Center, the diference between the two is that “top news aggregates the most interesting content that your friends are posting, while the most recent filter shows you all the actions your friends are making in real-time” (Facebook, 2011). Fast forward to 2015 and the Help Center describes the news feed in quite diferent terms. No longer is there a distinction made between a realtime feed and the “top news” feed. Gone are terms such as aggregation and filter or any other explicit trace of the software doing something. According to Facebook as of January 2018:

74

I F . . . T H E N

Posts that you see in your News Feed are meant to keep you connected to the people and things that you care about, starting with your friends and family. Posts that you see first are influenced by your connections and activity on Facebook. The number of comments, likes, and reactions a post receives and what kind of story it is (example: photo, video, status update) can also make it more likely to appear higher up in your News Feed. (Facebook, 2018)

What we see is the apparent disappearance of the algorithm, seemingly replaced by the agency of the user and the network. Recall just how the logic of the news feed was described by the Help Center in late 2013: “The News Feed algorithm uses several factors to determine top stories, including the number of comments, who posted the story and what type of post it is” (Facebook, 2013). In the midst of what became publicly known as the Facebook emotion experiment during summer 2014, Facebook’s Help Center conveniently replaced any mention of the algorithm with human-centered factors. The changes that were made to the Help Center descriptions of the news feed back in 2014 still more or less persist to this day. Now the regime of visibility on the news feed is seemingly determined by “your connections and activity” and “the number of comments and likes a post receives.” If users were feeling that they were being experimented on by having their news feeds “manipulated,” the change in description may arguably be seen as Facebook’s atempt to give back some level of control—but also responsibility—to users, if only rhetorically. But we are geting ahead of ourselves.

Let me quickly revisit some of the important changes and developments to the news feed and its algorithmic logic in order to provide some context to the arguments made. Many of Facebook’s most important features and functionalities were implemented in the years 2009–2011, including the “like” buton, the open graph protocol, messenger, and timeline. For users, the changing news feed has always been an issue of heated debate and anguish, especially in terms of the many changes in default setings, privacy setings, transparency, and the usability of the platform. In earlier years, the tabs distinguishing between diferent versions of the feed were visually diferentiated to a greater extent and, therefore, easier to control for users. Or so it seems. While one would think that having two feeds called “most recent” and “top news” would imply a clear distinction between an unfiltered and a filtered feed, respectively, a decade of Facebook history has repeatedly shown that nothing can be taken for granted. During February 2011, Facebook changed the setings and options for the “most recent” feed. Two basic setings were implemented, a seemingly unfiltered one, showing stories from “all of your friends and pages” as well as a filtered one displaying only “friends and pages you interact with most.” It was not so much the fact that Facebook changed the news feed feature yet again that brought about public outcry on the mater but the fact that Facebook had changed the default seting of the “most recent” feed to “friends and pages you interact with most”

Life at the Top

75

so that the feed most users believed to represent every update from every friend in a real-time stream, in fact, became an edited one, much like “top news.” What caused the most controversy was the fact that this change in default occurred without Facebook notifying its users about it. The option to change the default was tucked away at the botom of a drop down menu, barely noticeable. Users who did notice, however, quickly started to warn other users, pointing to the fact that these changes basically meant that people were not seeing everything that they should be seeing (Hull, 2011).

The revelation that the “most recent” feed was filtered is just one of many controversies surrounding the features and functionalities of the Facebook platform that the company has faced throughout the years. Besides representing yet another case of shaky privacy setings, it clearly points toward a certain disconnect between user expectations and reality. As we will discuss in much more detail in the next chapter, user expectations and perceptions about how technology works or ought to work may afect their media practices just as much as the actual working of that technology. As with the example above of users geting a sense of not seeing everything they should be seeing, the aforementioned revelations about the Facebook emotion experiment continue to show how important it is to hold software companies such as Facebook accountable. In late June 2014, the news broke that Facebook had “tinkered” with users’ emotions by manipulating the news feeds of over half a million randomly selected users, changing the number of positive and negative posts they saw (Goel, 2014). Facebook researchers had published the findings from the experiment in an academic paper. The experiment itself had been carried out for one week in January 2012, one and a half years prior to the public outcry engendered through the mass media. Notwithstanding the apparently small efects obtained, the experiment unleashed a huge controversy about algorithms, human subject research, and regulation (Meyer, 2014). What was most surprising about the media frenzy following the academic publication was not so much the actual experiments described in the article but the widespread public surprise in the face of this news. Like the public perception of what a “most recent” feed ought to be, the Facebook emotion experiment revealed that there are many persistent beliefs about how technology works and should work. As Gillespie (2014) put it in a blog post:

There certainly are many, many Facebook users who still don’t know they’re receiving a curated subset of their friends’ posts, despite the fact that this has been true, and “known,” for some time. But it’s more than that. Many users know that they get some subset of their friends’ posts, but don’t understand the criteria at work. Many know, but do not think about it much as they use Facebook in any particular moment. Many know, and think they understand the criteria, but are mistaken. Just because we live with Facebook’s algorithm doesn’t mean we fully understand it.

76

I F . . . T H E N

Indeed, we are only at the beginning of figuring out what it means to live in a society increasingly mediated and governed by algorithms. While it has been “known” for some time that Facebook uses algorithms to select the stories shown at top of the news feed, it is important not to forget how recent this time span really is. Moreover, what exactly this knowledge amounts to varies greatly. While we might never know for sure how the technology—the software and algorithms—of Facebook work in detail, there are enough ways in which we might know its operational principles and logics. As I argued in the previous chapter, algorithms are never so black or so boxed that they do not allow for critical scrutiny of their functions, assumptions, and embedded values. The cybernetic lead does not ask us to reveal the exact content of the box but to experiment with its inputs and outputs. In the following, I will walk you through some of the most important criteria the Facebook algorithm uses to serve people the “most interesting” content on their news feeds. The materials used to present some of the operational logic of Facebook’s algorithmic workings I draw from the technical specifications of specific inventions that pertain to the news feed as disclosed and discussed in white papers, computer science conference papers, Facebook patent applications, and recorded talks given by a Facebook engineer. Moreover, I draw on various news reports and blog posts from the technology press, most notably from publications such as Wired, The Atlantic, Techcrunch, Mashable, the New York Times, and the Guardian. These documents have been read and coded for the technological mechanisms entailed by the news feed. Technical specifications entailed in patent applications, in particular, contain much important information about how these systems work and are imagined by their inventors. Like any piece of discourse, patent applications also have to be read carefully and within the realms of what they are produced to achieve as rhetorical devices in a particular commercial context. To test some of the operational logics described in these documents, I also report on a technography of my own news feed. What, then, are the principles and logics of Facebook’s algorithmic form of editing the news feed?

The News Feed Ranking Algorithm

At Facebook’s third f8 developers’ conference in April 2010, Facebook for the first time “spilled the beans of some of the magic that is the news feed” as engineer Ruchi Sanghvi put it, introducing developers (and the public) to what was then referred to as the EdgeRank algorithm.11 Today, the notion of EdgeRank has largely been replaced by the much more generic notion of the Facebook algorithm(s) or the news feed ranking algorithm, respectively. Here, I use these terms somewhat interchangeably, sometimes referring to EdgeRank but, most of the time, simply referring to the ranking algorithm governing the news feed as Facebook algorithm.12 Let us take it from the start. When introducing the concept of EdgeRank in 2010, engineer Ari

Life at the Top

77

Steinberg explained it as an algorithm “basically looking at the edges in the graph” to decide how to rank the stories in news feed.

For Facebook, every item (i.e., status update, image, video) in the graph is considered an object. Every interaction with the object—for instance, through a “like” or a “comment”—creates an “edge.” The formula suggests that the rank of an object in the graph is the sum of the ranks of the individual edges. To paraphrase Ari Steinberg, one may think of each edge as having three main components: a time when it happened, a type of edge (a comment or like), and a user who created that edge. For each user, Facebook computes a score on how much a particular viewer tends to interact with the user—for instance, how often they comment on the user’s updates, look at their profile, or send messages. The edge weight component of the formula assigns a weight to each type of edge; for example, a comment may weigh more heavily than a like and so on. Steinberg also suggests that they train models to figure out what the appropriate weights are for diferent types of edges, hinting at the use of machine learning techniques. The last component is time decay because, as Steinberg suggests, freshness tends to be prety important (Kincaid, 2010). EdgeRank is calculated by multiplying the afinity, weight, and time decay scores for each edge. Nowadays, however, the Facebook algorithm cannot simply be broken down into three distinct elements (if it ever could). Any time a user visits Facebook, there are on average 1,500 stories for them to see. Needless to say, only a few of these stories can be shown at any given time. The task for Facebook, then, is to decide which stories are the most deserving. By looking at friend relationships, frequency of interactions, number of likes and shares a post receives, how much a user has interacted with particular types of posts in the past etc. Facebook is “capable of assigning any given Facebook post a ‘relevancy score’ specific to any given Facebook user” (Oremus,  2016).13 Once the relevancy score for each post has been determined, a sorting algorithm can then put them in the right order for the user to see.

Becoming visible on the news feed—appearing in that semi-public space— depends on a set of inscribed assumptions on what constitutes relevant or newsworthy stories. How many friends comment on a certain piece of content, who posted the content, and what type of content it is (e.g., photo, video, or status update) are just some of the factors at work in determining the rank of an edge. The higher the rank, the more likely it will be that an object appears in the user’s feed (Kincaid, 2010). As explained in a Facebook patent application, the system “uses diferent types of information about users in the process of selectively providing content to users, including user profile objects, an action log, and edge objects” (Luu, 2013). User profile objects refer to all the declarative information that users have explicitly shared with the system. The action log stores diferent kinds of interactions taking place on the platform—for example, becoming a fan of a musician or adding an event. Edge objects store information about users’ connections to other nodes, including the strength of the connections between users, which the system computes into an afinity score (Luu,  2013). In other words, the algorithm is based on the assumption that users are not equally connected to their friends. Some friends “count more” than others. The friends that count more are those with whom a user interacts on a frequent basis or a more “intimate” level—say, by communicating with a friend via “chat” rather than on the “wall.” The news feed algorithm is also geared toward highlighting certain types of edges while downgrading others in which the type of interaction becomes a decisive factor. Chating with someone on “Facebook Chat” presumably counts more than “liking” his or her post. Through tests, Facebook found that, “when people see more text status updates on Facebook they write more status updates themselves”—particularly, when they were writen by friends as opposed to pages (Turitzin, 2014). As product manager for news feed ranking Chris Turitzin (2014) says, “Because of this, we showed people more text status updates in their news feed.” Clearly, the type of content that is more likely to generate an interaction or prompt users to take action is made more visible and given priority. As Facebook engineers Andrew Bosworth—who is often cited as the inventor of news feed—and Chris Cox wrote in a patent document, the system determines an overall afinity for past, present, and future content based on one or more user activities and relationships:

78

IF . . . THEN

Some user interactions may be assigned a higher weight and/or rating than other user interactions, which may afect the overall afinity. For example, if a user emails another user, the weight or the rating for the interaction may be higher than if the user simply looks at the profile page for the other user. (Bosworth and Cox, 2013: 6)

There is a certain circular logic embedded in the algorithm. In order for you to like or comment on a friend’s photo or status update, they have to be visible to you in the first place. Any time a user interacts with an edge, it increases his or her afinity toward the edge creator. For instance, we can assume that comments outweigh “likes” as they require more individual efort. The system assigns ratings and weights to the activities performed by the user and the relationships associated with the activities. Many diferent variables may be used to assign weight to edges, including the time since the information was accessed, the frequency of access, the relationship to the person about which information was accessed, the relationship to a person sharing common interests in the information accessed, the relationship with the actual information accessed (Bosworth and Cox, 2013: 5). In deciding what stories to show users first, Facebook also prioritizes certain types of content over others. For example, content items related to so-called life events such as a weddings, the birth of a child, geting a new job, or moving across the country may be “prioritized in a ranking of news feed stories selectively provided to users to ensure that the most relevant information is consumed first” (Luu, 2013: 2). The weight given to certain types of edges, moreover, is likely to depend on the internal incentives that Facebook may have at any given point in time. If the objective for Facebook is to promote a certain product—for instance, the “questions” or “places” feature, interactions with these features will probably be ranked higher than others. This is  understandable. News feed is the best way to increase awareness of new (or neglected) features and functionalities.

Life at the Top

79