If...Then Algorithmic Power and Politics

This brings me to a third diference between November two years ago and today. Socially and culturally speaking, algorithms have emerged as a social and cultural imaginary of sorts. Algorithms are seemingly “everywhere.” They have become sites for cultural and social production. As objects of news stories, academic papers, and conferences, as well as the focal point of public controversies, popular discourse, cultural production, and afective encounters, algorithms are producing calculative results. And while algorithms are oriented toward us, we, too, are increasingly becoming oriented toward the algorithm. In chapter 5, I showed how the algorithmic output of social media becomes culturally meaningful, as seen in the ways that people form opinions about specific systems and act strategically around them.2 Algorithms are not just making their mark on culture and society; to a certain extent they have become culture. As Roberge and Melançon suggest “algorithms do cultural things, and they are increasingly active in producing meaning and interpreting culture” (2017: 318).

However, algorithms are not simply means of interpreting culture, they are also productive of culture, understood in terms of the practices they engender. Hallinan and Striphas claim “engineers now speak with unprecedented authority on the subject, suffusing culture with assumptions, agendas, and understandings consistent with their disciplines” (2016: 119). Although engineers and computer scientists are perceived to be in a privileged position to hold opinions about the term “algorithm,” we are at the same time seeing an emerging trend where “ordinary” people and institutions are speaking and thinking about algorithms. They do so by permeating algorithms with assumptions, agendas, and understandings that are consistent with their specific backgrounds and life worlds. For someone who encounters the Facebook algorithm as a suppressor and regulator of content, the algorithm gets sufused with agendas and assumptions that may not be shared by someone else who encounters similar types of posts but in a diferent situation.

150

I F . . . T H E N

What, then, are we to make of the algorithm amid the condition of perpetual change? How can we know algorithms and the systems they are part of, if they are constantly changing? This book has suggested that neither the changing nature of the object nor the way it is engulfed in trade secrets and proprietary systems should be seen as a necessary deterrence. I have argued that our understanding of how algorithms shape everyday life is not inhibited by not knowing the thing, simply because the thing is always already “many quite diferent things” (Law, 2002: 15). In chapter 3 I ofered three methodological tactics one could apply in the analysis of algorithms. Each of these maps to how algorithms were examined in chapters 4, 5, and 6, respectively. One approach serves as an aid in delineating how technologies materialize cultural imaginaries. Here, the notion of technography was presented as a way of interrogating algorithms as spatial and material arrangements that help to configure users in certain ways (chapter 4). Second, I proposed a phenomenological understanding of the productive force that people’s imaginaries of algorithms have for the practical engagement with their lived-in environment (chapter 5). The third (but by no means final) option was to think the material and the discursive together, by looking at the ways in which algorithms are diferently configured and made to mater in specific setings (chapter 6).

Developing an understanding of the power and politics of algorithmic media depends on acknowledging their “variable ontology.” As Introna and Wood contend, “we cannot with any degree of certainty separate the purely social from the purely technical, cause from efect, designer from user, winners from losers, and so on” (2004: 180). To counter the assumption that algorithms are essentially technological things or coded objects that can be easily delineated as separate entities, in this book I sought to examine what I have called the ontological politics of algorithms. Based on the manner in which notions of the machine, software, and artificial intelligence are bound up in and entangled with sociality, subjectivity, and social issues, the ontological politics of algorithms refers to the ways in which these categories are made diferentially relevant and available as part of specific configurations. The manyfoldedness of algorithms, moreover, raises the question of whether a distinction between knowing subjects and objects known is misguided to begin with. Perhaps, as Mol suggests, we are beter served by “spreading the activity of knowing widely” (2002: 50). In chapter 3 I therefore suggested a particular way of atending to the variable ontology of algorithms by shifting atention away from questions of ontology—of where agency is located and whom it most obviously belongs to—to questions of politics, of when agency is mobilized and to what end. In order to understand the power and politics of algorithms, there is a need to atend to the question of “when the algorithm is”?

The “when” of algorithms becomes particularly evident if we look at some of the recent public discussions concerning algorithmic systems that are perceived to have gone astray. Such public controversies reveal how algorithms are often materialized diferently depending on the seting. In the case of Facebook’s “trending topic controversy,” which hit the news in May 2016, human editors were accused of censoring conservative topics on the platform’s trending topic feature. The story was largely framed in terms of “biased” and “subjective” humans distorting the “neutral” and “objective” algorithms responsible for ranking and sorting Facebook’s trending topic feature. Interestingly, only a couple of months later, in early September 2016, Facebook became the target of another “algorithm controversy.” Yet this time, the algorithm was accused of doing the censoring. Several prominent members of the public, the then Norwegian prime minister, Erna Solberg, among them, expressed worries about the fact that Facebook’s news feed algorithms had deleted a famous Pulitzer Prize winning war photograph from people’s news feeds. As described in the previous chapter, the story culminated when the editor-in-chief of Norway’s national newspaper Aftenposten challenged (with great pathos) Mark Zuckerberg to take editorial responsibility for the algorithm’s actions. These controversies show how algorithms can be mobilized and materialized in diferent manners, depending on the specific constellations they are part of. Also, they demonstrate that we cannot with any degree of certainty separate the purely social from the purely technical—nor should we necessarily want to. In both of these cases there was an algorithmic system at play, including both humans and machines. If we want to understand questions of power and politics, then, it is not enough to look at one element only. Precisely because we are not dealing with two situations of radical diference but, rather, with situations were these diferences were made, an understanding of algorithmic power and politics must necessarily entail an examination of such boundary-making practices.

Algorithmic Life

151

When algorithmic systems “do wrong,” or are otherwise perceived as erroneous, questions frequently arise as to “who is to blame,” “who or what is responsible,” and “whose fault it was, the algorithm or the human.” While it might not be the most satisfactory of answers, at least not from a tabloid standpoint, this book has made clear how the answer must always remain partial. Even more important is the claim that, by framing questions as an opposition of terms, as it often done in the public discourse surrounding algorithms, the risk is to simply think of algorithms as purely technical, or mechanical entities devoid of any human agency. When the Microsoft bot Tay started to tweet obscene and misogynist statements, or when Google image search suggested “ape” as an appropriate tag for a person of darker skin color, “the algorithm did it” is only an acceptable response if by algorithm we mean the “sociotechnical assemblage that includes algorithm, model, target goal, data, training data, application, hardware—and connect it all to a broader social endeavor” (Gillespie, 2014). A core tenet of this book is therefore to think of algorithms not merely as “coded instructions telling the machine what to do,” but rather as emergent accomplishments or socio-material practices. How algorithms come to mater in contemporary society is not about trying to define what they are or at what points they act, but rather about questioning the ways in which they are enacted, and come together to make diferent versions of reality.

152

I F . . . T H E N

Technicity

How are we to think of the capacity that algorithms have to bring new realities into being? At the start of the book, I suggested a specific approach to the question of “how algorithms shape life in the contemporary media landscape,” by looking at how platforms like Facebook condition and support specific forms of sociality in ways that are specific to the architecture and material substrate of the medium in question. I described how Facebook encodes friendship in ways that essentially support the circuit of profit, and introduced the notion of programmed sociality. At the same time, in chapter 1 I warned against taking this notion of programmed sociality to credit for technological determinism. Software and algorithms do not simply operate in isolation or exercise power in any unidirectional way. Rather, their capacity to produce sociality always already occurs in relation to other elements, and as part of an assemblage through which these elements take on their meaning in the first place.

This is the technicity of algorithms, understood as the capacity they have to make things happen as part of a co-constitutive milieu of relations. For the French philosopher Gilbert Simondon, who introduced the concepts of transduction and technicity as useful frames to account for the productive power of technical objects, humans and machines are mutually related, where the technical object always already appears as a “theatre of a number of relationships of reciprocal causality” (1980: 22). Technicity, then, pertains not so much to what the technical objects are, but rather to the “forces that it exercises on other beings as well as in and through the new virtualities, and hence realities, it brings into being” (Hoel & van der Tuin, 2013: 190). This understanding comes close to the Foucauldian manner in which power was perceived in chapter 2. For Foucault, power pertains precisely to its exercise and force, to the “way in which certain actions modify others” (1982: 788). For an understanding of algorithms, this implies resisting reading algorithms as either technology or culture. Rather one is prompted to think of algorithms as “dramas”—or “theatre” of relationships in Simondon’s terms—that algorithmic mediations bring about. The concept of technicity, then, ofers a way of thinking about the productive power of algorithms that doesn’t rely on the atribution of power to some fixed characteristics or stable artefact devoid of human agency. To speak of the technicity of algorithms is to emphasize how algorithms do not possess power in and of themselves, but how power unfolds as an “ontological force” (Hoel & van der Tuin, 2013).

Let’s briefly consider the examples of Amazon book recommendations and Twiter’s “While you were away” feature. The technicity here should be understood as the coevolving conditions that emerge from the algorithm’s dynamic functioning. An analysis of an event such as the recommendation of a book, or having one’s fear of missing out catered to by a promise of a peek into what happened “when you were away,” should be grounded in an “understanding of the nature of machines, of their mutual relationships and their relationships with man, and of the values involved in these relationships” (Simondon, 1980: 6). Though I have not made any claims about “the nature” of algorithms, I propose that algorithmic modes of existence are grounded in the relational drama of diverse environments. In the process of recommending books, Amazon’s “collaborative filtering” algorithms create conditions through which certain books are brought to user atention while others remain elusive. Collaborative filtering is based on the assumption that customers who share some preferences would also share others. Rather than matching users with similar customers, Amazon uses “itemto-item” collaborative filtering to recommend similar books or other items, and this calculated chain of similarity relies heavily on user input. Thinking of the power of algorithms in terms of their technicity leaves no one of the agential hook, so to speak. The nature of the machine always needs to be understood in relation to other machines and “man,” in Simondon’s terms. On Amazon, we buy books (often several at the same time), put them into wish lists, search for items, browse the platform, write reviews, and provide ratings. On Twiter, we tweet, browse, spend time, build nonreciprocal networks of relationships, reply, and retweet. We are already implicated. According to Barad, “we are responsible for the world in which we live not because it is an arbitrary construction of our choosing, but because it is sedimented out of particular practices that we have a role in shaping” (2007: 203). This means that we cannot treat the power of algorithms as a one-way street. If power refers to the way in which certain actions modify others, our actions count too. The tweets we post, items we purchase, and likes we hand out all factor in to the sum total. As we saw in chapter 4, actions such as these modify what becomes visible to us and to our networks.

Algorithmic Life

153

This also means that the roles of designers, developers, and other decision-makers are crucial. Just because algorithms learn and amplify existing societal biases, inequalities or stereotypes, it does not mean that they cannot be corrected for. On the contrary, algorithm developers can compensate for the bias in datasets, and companies do make choices about when to intervene in correcting certain algorithmic outcomes.3 When algorithms end up as biased toward certain groups, these instances can be identified and handled computationally. As was discussed in chapter 6, using algorithms in journalism does not automatically entail the creation of filter bubbles, contrary to what is often claimed. As one of the interviewees suggested, “If you want your algorithm to reflect diversity that’s what you have to design for.” What’s important is how designers, developers and decision makers think about what algorithms should be optimized for, and what possible consequences this may have for diferent social groups.

Orientations

Agency is not all or nothing. To say “the algorithm did it” will not do. Nor is it an option to assign responsibility to others, just because there are always other entities involved. As with the controversies surrounding algorithms discussed in the book, it is not a question whether algorithms conceived of as non-human objects or humans did this or that, but how the atribution of agency is realized in particular setings. Yet, the question of “how exactly to be responsible, how to hold or to be held accountable if agency is distributed?” (Simon, 2015: 153) remains. As crucial as this question is, distributed agency does not make accountability impossible. On the contrary, the agential messiness entailed by such a relational view makes accountability mandatory (Barad, 1998). There are at least two aspects to this. First, agency is not necessarily equally or symmetrically distributed. Secondly, the process of making certain elements more or less available or unavailable to specific actors is not something that occurs naturally, but rather enacted through boundary-making practices. What does this entail for an understanding algorithmic power and politics?

154

I F . . . T H E N

First, it needs to be pointed out that while agency, in theory, is symmetrically distributed, in practice this is not necessarily the case. Diferent entities do not hold agency or exercise power in the same way.4 When Foucault stated that “power is everywhere,” he did not suggest that everybody has equal access or opportunity to exercise power (1990: 93). Rather, power is “the moving substrate of force relations which, by virtue of their inequality, constantly engenders states of power, but the later are always local and unstable” (1990: 93). Because power is not something that one group, person, or thing holds, the analyst’s task is not about identifying who holds power or where the proper source of action is, but to understand “the efects of particular assemblages, and assessing the distributions, for beter and worse, that they engender” (Suchman, 2007: 11). In order to understand how algorithms exercise power and what it would mean to hold them accountable, we need to assess the distributions that algorithmic forms of operations help to engender. In chapter 4, the operative functioning of algorithms was examined with regard to the forms that “categorize the individual, marks him by his own individuality, atach him to his own identity [. . .] a form of power which makes individuals subjects” (Foucault, 1982: 212).

It maters how algorithmic systems position subjects. In going forward it is therefore crucial that we develop a beter understanding of how we are addressed and positioned as objects of algorithmic atention. What sorts of subjects become possible in the algorithmic landscape? The research presented in chapter  4 on the Facebook news feed suggests that the algorithms work to enact and support a form of participatory subjectivity that hinges on continued and ongoing engagement with the platform. Tellingly, a recent patent document filed by Facebook expresses this very sentiment: “Stories are frequently selected for inclusion in a newsfeed because they are likely to be of interest to the user viewing the newsfeed. This encourages the user viewing the newsfeed to continue using the social networking system” (Rubinstein et al., 2016). While the particular configurations of the Facebook algorithms have certainly changed since chapter 4 was writen, the basic factors guiding the ranking of the news feed remain. These are afinity (i.e., the strength and nature of interactions between friends), weight assigned to types of objects (i.e., status updates, videos, comments), and time. In general terms, algorithms do not control what users say or how they behave. Rather, algorithms shape how users come to speak and what actions are made possible to begin with.5 In the terms of Foucault, algorithms can be seen as instantiating a particular form of “government,” in terms of “the way in which the conduct of individuals or of groups might be directed [. . .] To govern, in this sense, is to structure the possible field of action of others” (Foucault, 1982: 790).

Algorithmic Life

155

While algorithms do not determine how people behave, they shape an environment in which certain subject positions are made more real and available to us. As Sara Ahmed suggests in her book Queer Phenomenology, we are “directed in some ways more than others” (2006: 15). Her book takes up the question of what it means to be oriented, that is, to become atuned in certain ways. Ahmed instructively asks: “What diference does it make ‘what’ we are oriented toward?” (2006: 1). While Ahmed talks about human orientations mainly in terms of how social relations are arranged spatially, algorithms are certainly also oriented toward users. Algorithms, particularly in the age of machine learning, need us, depend on us, and thrive on us. In the midst of ongoing, fervent arguments about Facebook disseminating “fake news,” Facebook worries about “data-dirtying” practices, whereby users willingly provide misinformation (Ctrl-Shift, 2015). As chapter 5 shows, these worries are far from unwarranted. People do not necessarily post what is on their mind (to paraphrase Facebook’s status update prompt) or what is actually happening (to use Twiter’s expression). Instead, people may post much more strategically, in atempts to make themselves more or less “algorithmically recognizable” (Gillespie, 2017).

Ahmed suggests that “orientations mater,” because they shape how the world coheres around us. That is, orientations mater because this is “how certain things come to be significant” as part of situated encounters (Ahmed,  2010: 235). Moreover, Ahmed writes, orientations mater in that it “afects how subjects and objects materialize and come to take shape in the way that they do” (ibid.). As I have suggested, algorithms are not just oriented toward us, we are increasingly becoming oriented toward algorithms and algorithmic systems as well. In chapter 5 I show how we are developing our own sense of sociality and sense of self in, through, and around algorithms. There is an interesting relational drama at play when a single person gets served an ad for engagement rings or a middle-aged woman gets suggested recommendations for wrinkle cream. Was it an algorithm or merely the person’s clicking behavior that was responsible for these recommendations popping up, and to what degree does it mater? Can one simply atribute agency to the machine, and laugh of the stereotypical ways of depicting personal relations inscribed into the design of algorithms? What would it mean to take responsibility in this case?

I suggest in chapter 5 that the atribution of agency is realized in the encounters that people have with what they perceive as algorithmic phenomena. Depending on specific situated encounters, algorithms are perceived as creepy, helpful, disturbing, intrusive, and so on. The notion of the “algorithmic imaginary” was introduced to denote ways of thinking about what algorithms are, what they should be, how they function, and what these imaginations, in turn, make possible. While algorithms may produce certain conditions for action, these conditions are not necessarily atributable to algorithms in a purely technical sense. Rather, how and when people perceive algorithms may sometimes mater even more. Thus, when looking at what people do online, there is no way of telling what made them act in certain ways. The notion of the algorithmic imaginary suggests that what the algorithm is may not always be about the specific instructions telling the computer what to do, but about the imaginations and perceptions that people form. It follows from the logic of machine learning, that what people “do in anticipation of algorithms,” as Gillespie suggests, “tells us a great deal about what algorithms do in return” (2017: 75). In terms of accountability and responsibility, then, we are all implicated, machines and humans alike.

156

I F . . . T H E N

Boundary-making Practices

From a strictly technical perspective, an algorithm can be defined as a step-by-step procedure for solving a problem in a finite number of steps. Throughout this book, I have argued that algorithms are much more besides this narrow, technical concept. Algorithms do not only instruct, they mean something, and often they mean diferent and conflicting things. Algorithms manifest as objects of social concern, as they become inscribed into the fabric of everyday life. Algorithms, moreover, help to shape the ways in which we come to know others and ourselves.

If the textbook definition of algorithm is but one version implicated in the manyfoldedness of the term, then where and when besides code do algorithms manifest? As I atest in this book, algorithms exist in the conversations of academics, in media portrayals, public controversies (i.e., debates over “fake news” on Facebook), in people’s perceptions and imaginations, in representational resting points (i.e., simplified images of formulas), in metaphors (i.e., the notion of algorithm as recipe), as part of films and popular imagery, stories, and professional practices. In an expanded understanding of the term, algorithms are seen as events. Without front-loading an ontological commitment with regard to algorithms, the notion of the event helps directing atention to the ways in which algorithms make other things suddenly conceivable.

Who or what is made accountable and responsible for algorithmic outcomes depends on how and where the boundaries are drawn. Disentangling the ontological politics of algorithms requires “remembering that boundaries between humans and machines are not naturally given but constructed, in particular historical ways and with particular social and material consequences” (Suchman,  2007: 11). Again, there are at least two sides to this. What counts as an algorithm, as well as when it counts, are not given. Following Suchman, we need to pay atention to the “boundary work through which a given entity is delineated as such” (2007: 283). As we have seen, what counts as an algorithm may vary greatly. Yet, most people would probably have litle problem acknowledging that the computer science textbook definition of an algorithm, at least partly, constitutes the phenomenon in question. With this book, I hope to have made people more susceptible to the idea that algorithms exist and manifest on scales and levels that go far beyond the narrow, traditional definition.

Algorithmic Life

157

Clearly, the question of what counts as an algorithm is highly site-specific, and even when algorithms are assumed they may count diferently. In terms of professional practice the interviews presented in chapter 6 reveal the extent to which algorithms perpetuate designer’s values, beliefs, and assumptions. I also showed that they reflect the diferent institutional and organizational setings in which algorithms linger. News media professionals frequently contrast the “good” journalistic algorithm to the “inferior” algorithms of social media platforms. News editors and managers speak of the importance of coding the “right” journalistic values into systems, equating the “bad” algorithm with systems that merely give people more of what they already had. While it might be tempting to take sides, to say that one algorithmic system is beter than the other, we must also be wary not just of the politics embedded in design but the politics of making boundaries. As was suggested in chapter  6, what constitutes a “problematic algorithm” in one seting (i.e., the Facebook algorithm challenging the livelihood of journalism), may simply be rendered unproblematic and desirable in another (i.e., Facebook itself touting the algorithms as merely helping people get closer to their friends). Problematic values in design aside, it is also important to point out that there are no “innocent” ways of knowing or talking about algorithms.

Returning to what was the starting point of the book, we are now in a beter position to think through some of the more general claims that were made at the outset. The question was raised as to how algorithms are shaping the conditions of everyday life. One answer is that the materiality of algorithms, the values and assumptions designed into the technological properties of algorithmic systems, govern sociality in specific ways. A claim was made to the efect that algorithms “program” ways of being together, producing the conditions through which people come to speak and connect online. Now we are hopefully in a beter position to see how questioning the ways in which algorithms shape life also requires us to question how life shapes algorithms. Through the continuous collection of user data, algorithmic systems reach a higher level of flexibility and responsiveness: The relations between self and others change continuously. What I want to suggest, then, is that the notion of programmed sociality at play in the algorithmic media landscape is one where what’s at stake are the ongoing actualization of “becoming together” in ways that temporarily stabilize, rather than a result of pre-programmed forms of being. The way we are together can only be thought of in terms of becoming, where the “we” implied by sociality is never confined to any specifies or essence, but to the “we” as relation.

158

I F . . . T H E N

Programmed sociality, then, implies a mutual reciprocity between machine and human, and the recognition that however life is shaped, responsibility is shared.6

The “if. . . then” conditional statements of algorithmic rules imply that certain conditions must be met in order to have particular consequences. But in order for algorithms to solve problems and have consequences, they are contingent on a wide variety of people, technologies, investments, and resources (Neyland & Möllers, 2017). Just as algorithms and platforms are perpetually changing, so are the worlds created in and around them. In other words, as the “if ” changes, so does the “then.” The way in which one can solve the problem of showing the “right content, to the right people, at the right time” is never set in stone. As the content changes, whom the content is right for might change as well. The right time for one person may be uterly wrong for someone else. When algorithms draw boundaries by deciding which of the 1,500 potential Facebook stories to make visible on a particular user’s news feed, such decisions are never made in isolation of people. When people make boundaries by atributing algorithmic outcomes to inferences made by a piece of computer code, these atributions rely on the material arrangements that let people make such claims to begin with.

If the machine is involved, then people are too. The notion of algorithmic power and politics is not about the ways in which algorithms determine the social world. Nor is it about what algorithms do in and of themselves. Rather, we have to ask how and when diferent aspects of the algorithmic are made available or unavailable to actors in a given seting. In looking at what constitutes an algorithmic life, the central question therefore becomes: who or what gets to be part of whatever is being articulated as the algorithm.

Algorithmic Life

159