If...Then Algorithmic Power and Politics

METRIC POWER

As with any commercial business, what is of course greatly valued is profit. The al gorithmic potential is less about the specific code than it is about the economic promise that algorithms provide, and their perceived capacity to act on the data in new ways. If algorithms promise to reduce production and circulation costs, there might not be much to argue against. As Stig Jakobsen, editor-in-chief at the local Norwegian newspaper iTromsø, told me when I interviewed him in May 2016, the choice of implementing an algorithm to personalize the “mobile-front” of the newspaper was rather easy: “If you find something that gives you 28% more trafic, as we have done, all you can do is say thank you” (Interview 2, May 2016).11 We may call this the pragmatic route. Here quantification and the measurement of trafic is the new language of success, readily supported by algorithms that allow for seemingly objective decision making. Increased access to user data implies new insights and new possibilities for action. Jakobsen, who came to the local newspaper in mid-2015 after having spent much of his career in the magazine business as founder and editor-in-chief of some of the biggest Scandinavian lifestyle magazines, says real-time analytics and algorithms promise to support more informed journalistic decisions. In the magazine business as in journalism more generally, decisions were often taken purely on the basis of a “gut feeling.” In times of economic hardship, however, a “gut feeling” is too risky, while metrics and numbers can potentially save you many unnecessary missteps. Though Jakobsen is enthusiastic about the seemingly objective promise of algorithms to deliver more relevant news to the right readers, he is careful not ascribe too much power to algorithms as such. The algorithm is merely a tool that can be used to make sure the right news reaches the right person, Jakobsen suggests, “If you are not interested in Justin Bieber, the algorithm will not recommend that article for you” (Interview 2, May 2016).

The power of metrics to influence editorial decisions, however, is not new (Anderson,  2011; Domingo,  2008; Vu,  2014). As journalism scholar Chris W. Anderson (2011) describes, based on his ethnographic research in Philadelphia newsrooms, new measurement and web trafic tools have contributed to the emergence of a “generative, creative audience.” Anderson found that “Editors were clearly adjusting their understanding of ‘what counted’ as a good news story based on the quantified behavior of website readership” (2011: 562). For digital director at the Danish newspaper Jyllandsposten, Jens Nicolaisen, the extent to which big data and new computational tools provide actionable insights is one of the most promising avenues for the news business. As Nicolaisen sees it, the widespread use of analytics and metrics inside newsrooms has resulted in something of a power shift in journalism more broadly. Previously, the news was largely produced from the inside out, he says, without too much thought dedicated to what readers actually wanted to read. This lack of journalistic responsiveness to audience desires was most famously observed by Herbert Gans in his seminal work Deciding What’s News (1979). Gans notes that he was surprised to find how litle knowledge about the actual audience existed within newsrooms: “Although they had a vague image of the audience, they paid litle atention to it” (Gans 1979: 229). Today, however, Nicolaisen observes, value creation increasingly starts from the outside—with what users actually click on, what they say they are interested in, what they share and talk about in social media (Interview 3, February 2015).

130

I F . . . T H E N

Ingeborg Volan, then director of innovation and newsroom development at Norway’s oldest newspaper Adresseavisen, shares this sentiment, suggesting that the landscape of journalism has changed profoundly with regards to digital media. Traditional newspapers were constrained by their technological afordances: “You simply could not print personalized newspapers for every subscriber,” says Volan. These technological constraints are now largely gone, and with it, an obvious rationale for producing the same paper for everyone. After all, Volan says, “Not all of our readers are alike. People’s interests vary a lot, and they read diferent things” (Interview 4, August 2016). This does not mean, however, that Volan thinks journalism should simply cater to people’s tastes and likes. “I don’t think anyone would want that,” she says when I ask her to elaborate on the power of users and readers to shape the news. It is not that power has shifted from editors to users— or to algorithms and data analytics for that mater. Although algorithms and social media platforms certainly challenge the privileged position news media organizations have historically occupied, they also provide a renewed opportunity for reaching people. As Volan sees it, it’s not an either or, but rather a question of careful realignment.

BETTER WORKFLOWS

The interviews also suggest that the potential of algorithms is strongly linked to their role in organizing news work, particularly with regard to being perceived as strategic change agents. From a managerial point of view, algorithms ofer an opportunity for speeding up the production cycle, making journalistic practices more eficient and news more consumer-oriented. It has, for example, become a fairly common practice for journalists to track their performance by monitoring how well their articles are doing in terms of readership and distribution. At the Danish newspaper Politiken, journalists and front desk editors are equipped with dashboards that help monitor and analyze how well specific articles and stories are doing. There are a few staple software systems that most organizations tend to use, including Chartbeat for tracking user behavior on site in real time, and Comscore or Google analytics for analyzing those data. Often elements from these “of-the-shelf” products are combined and adapted to the specific needs and interests of the organization in custom-made software solutions. As Anders Emil Møller, then director of digital development at Politiken, explains, dashboards and analytics software are important working tools that provide journalists with a beter understanding of their work processes. It may even give them a sense of success and agency, given that they are able to beter understand when an article is doing particularly well or what they may have to change to increase performance. A system like Chartbeat is used to monitor how much time a reader spends on a particular article and where he/she might move next. In this sense, software systems and dashboards become an epistemic tool, enabling an understanding of journalists’ own performance, or the behavior and interests of readers. In addition, they are used to realize the expectations that exist at the managerial level about what constitutes good news work. When asked whether the many diferent metrics and tracking mechanisms might possibly drain journalists in their work, Møller says these systems are merely there to help and support journalists to do the best they can (Interview 5, March 2014). Møller suggests that these systems may even support journalists in promoting their own work. For example, tracking systems are able to give journalists leverage to argue for a more prominent position on the front page, based on the metrics provided. In this sense, the dashboards and the metrics supplied by diferent analytic tools become imbued with discursive power used to negotiate with the front desk.

Programming the News

131

Yet, algorithms are primarily directed at the consumer, not so much at practitioners, Aske Johan Koppel Stræde, new director of digital development at Politiken says, when I interview him in September 2016. This is where Stræde locates the algorithmic potential, in helping journalists find, contextualize and frame the news in innovative ways. Rather than seting the agenda, algorithms can be used to provide journalists with new tools to “make it easier to set the agenda,” Stræde suggests. Whether these tools mine Twiter for important topics or track what people are talking about on Facebook, algorithms can be used productively to support journalistic practices, many of the informants suggest. During the interviews, I was struck by the way it appeared easier to talk about algorithms in a positive spin when they were invoked as human helpers, rather than as automated gatekeepers of sorts. Understood as less confronting, invoking algorithms as tools for making the editorial workflow more eficient can be a way of saying yes to algorithms without compromising on the democratic function of journalism. As we will see in the next section, the conversations around algorithms tend to shift when they are more directly linked to the values and institutional role of journalism. Journalists often see themselves as using algorithms to serve publics with beter news, and the managers tend to defend this public function of journalism as well.

While algorithms are variously regarded as helpful tools that support news workers in accomplishing their tasks and make the production cycle more eficient, they are not merely supporting journalists but also slowly changing the very condition of news work itself. Algorithms help to make journalism in new ways, by creating new genres, practices, and understandings of what news and news work is, and what they ought to be. As described in the burgeoning literature on digital and algorithmic journalism, algorithms are now used to find relevant stories on social media platforms (Thurman et al., 2016), create catchy headlines (van Dalen, 2012), use data sets to generate stories (Karlsen and Stavelin, 2014), and aggregate and recommend news stories ( Just & Latzer, 2017). In many of the news organizations studied large investments go into experimenting with new computational tools to make news work more efective. At Schibsted, Cramer explains, one problem is that journalists often write in a form that is too long, too complicated, and too intricate without really taking the user’s perspective into consideration (Interview 1, August 2016). There’s potential, as Cramer sees it, in the way algorithms can be used to “atomize” the news. Such an approach, which is also known as “structured journalism,” works by treating every bit of content as a separate block of data that can be reassembled into a new form, depending on what a specific reader already knows.12 As Cramer points out, by simply treating every sentence, word, or paragraph as a separate “atom,” newsrooms may use the power of algorithms and databases to produce news stories “on the go” and avoid unnecessary overlaps. Perhaps it is time to rethink the genre of the traditional news article itself, Cramer speculates, referring to a study conducted by Norway’s leading tabloid and most read newspaper VG, which showed an overwhelming overlap in content produced between their journalists on a daily basis. More than simply writing articles like journalists have always done, atomizing the news essentially implies a personalization of the storytelling process itself, a process that is fundamentally powered by algorithms and structured data. For Sundve, whose rhetorical mission at Schibsted seems in part to be about establishing “the true purpose of journalism” vis-à-vis technology companies, atomization, however, represents much more than a new computational technique. In the ongoing batle for atention, revenue, and control over informational flows, content is king. If what social media platforms do is merely to distribute content, creating original and personal content may just be the news media industry’s “greatest weapon in the fight for relevance” (Sundve, 2017). More important, perhaps, is the editorial accountability that comes with creating journalistic content. In order to succeed in keeping a distinct and accountable editorial voice, Sundve thinks it is crucial to develop new ways in which algorithms and people work together (Interview 7, May 2016).

132

I F . . . T H E N

To this end, it might be instructive to turn our atention to the Schibsted-owned Swedish mobile news app Omni as an example of the algorithmic potential found in organizing the editorial workflow in new ways. Relying on a combination of developers, editors, and a news-ranking algorithm, Omni seeks to provide users with what editor-in-chief Markus Gustafsson describes as a “totally new way of experiencing news” (Interview 8, April 2014). Gustafsson, who came from the position as managing editor at Sweden’s leading tabloid newspaper Aftonbladet before cofounding Omni with Ian Vännman, says it all started with a mission to create something like the Swedish version of Hufington Post. But instead of copying an existing news service, Omni developed into a unique service of its own. Rather than creating original content, Omni aggregates news from other sites and provides manually writen summaries of the news while linking to the original sources. Readers may also follow certain topics and personalize the mix of news. At the time of my visit to the Omni headquarters in Stockholm in March 2015, about ten editors worked in shifts to handpick the news. I spent a week on site, observing the editors day-to-day work, conducting informal interviews with some of them, and also conducting more formal interviews with designers, developers, and editors connected to Omni and the Schibsted media house (Interviews 9–16). The editors, who all come from a background as reporters and journalists before starting at Omni, do not write their own stories or work like reporters in a traditional sense. Instead, the work involves picking interesting news items from a content delivery system in order to make it “algorithm-ready.” That is, editors summarize existing stories using a custom-built content management system—for example, by giving the story a new headline, providing a short writen summary, linking it up to the original sources, and finding a suitable image to go with it.

Programming the News

133

A key part of making a story algorithm-ready (which, in this case, refers to preparing the news items in such a way that they can be processed and, subsequently, ranked by an algorithm) is seting news and lifetime values for each story. Before editors can publish the edited stories, they have to assign each and every one a  unique news value on a 5-point scale and a lifetime value on a 3-point scale. As Gustafsson explains, these values act as tokens of relevance and newsworthiness, and are subsequently fed into an algorithm that, ultimately, “decides exactly where and how to show the news” (Interview 8, April 2014). The algorithm ranks every story based on a combination of news and lifetime value, the reader’s personalized setings, time decay, and the relative weights predefined for each major news category (i.e., foreign news is weighted slightly higher than sports). The news and lifetime values that editors assign, Gustafsson suggests, are “subjective values” based on journalistic know-how and gut feeling (Interview 9, March 2015). The algorithm’s job, then, is to process the subjective values in an objective and predictable way. While the news-ranking algorithm is rather “banal and simple,” and the human efforts involved are “crucial and decisive,” Gustafsson is clear when he suggests that they “couldn’t have done it without the algorithm,” as “it does, in fact, influence quite a lot” (Interview 9, March 2015, and notes from field observations). There’s algorithmic potential in providing a beter user experience but, perhaps, “first and foremost in making the work flow more eficient and time-saving” (Interview 9, March 2015).

The (Un)Problematic Algorithm

While the notion of potentiality is an important aspect of how algorithms are talked about in journalism, algorithms are simultaneously conceptualized as more or less problematic. I say more or less because, when the algorithm appears as unproblematic it is largely due to its role being downplayed as trivial or insignificant. As with the notion of potentiality, algorithms become problematic in specific situations.

134

I F . . . T H E N

The interviews reveal how algorithms are deemed particularly problematic when they threaten to diminish the journalistic mission of ensuring an informed public. Moreover, algorithms become problematic when they threaten to automate too much—especially when they compromise people’s jobs, or when they become unruly and can no longer easily be controlled, starting to challenge the fundamental touchstones of editorial responsibility. Algorithms are not just problematic in terms of the various concerns and challenges to which they point: They are also conceived of as problematic at diferent times and circumstances. The notion of the problematic algorithm appears, for example, during moments of negotiation and breakdown, or when it becomes permeated with people’s worries and concerns.

Incorporating algorithms as agents in the newsroom is not without its challenges. Most news executives I talked to describe a variety of obstacles and trials encountered in the process of adding an algorithm to the mix. When algorithms are incorporated into the editorial workflow, the process is often characterized by much debate and tweaking in the beginning, before the algorithm gradually slips into the background and becomes, at least temporarily, stabilized. Many news outlets have dedicated teams of programmers and data journalists who experiment with new tools and techniques on an ongoing basis; however, when organizations instigate more pivotal and far-reaching technological changes, the process is often more top-down. A few dedicated people, often digital project managers and editors, make decisions that are subsequently rolled out into diferent sections and parts of the newsroom. First, you need to figure out what you want the algorithm to do. While some news organizations spend considerable time on concept development, others such as iTromsø, release the code to see what happens. Even in cases where algorithms are released more quickly, the process of implementation implies subjecting the algorithm to a series of trials and questions. As Jakobsen says of the process at iTromsø, “first, we had to ask ourselves, should we ask our readers for permission” and to what degree should the algorithm be opt-in or opt-out? (Interview 2, May 2016). When implementing an algorithmically generated newsfeed for the first time, they decided not to ask their readers for permission explicitly but rather to “just go ahead and ask for forgiveness later.” When I asked Jakobsen how often they think about or tweak the algorithm now, he replied, “not so much anymore.” In the beginning, “we tweaked it every day” until it became more stable and less problematic he says. While nobody seemed to complain about the opt-in solution, there were other issues that had to be dealt with, Jakobsen adds. It proved increasingly dificult to find the same article twice because the algorithm was constantly showing new items and personalizing content. To combat the disappearing archive that comes with personalization, iTromsø implemented a section called “my reading history,” where you could easily find old articles again. They also implemented an opt-out buton at the botom of the mobile interface to make it easier for readers to go back to a non-personalized feed. One of the things that is far more dificult to handle algorithmically, however, is knowing when to show certain content. In seting the parameters for personalization, having a sense of “right time” seems crucial (see also chapter 4). Like the annoyingly reappearing Lisbon hotel ads that plagued my Facebook news feed in the introductory chapter, Jakobsen uses a similar example to explain how an algorithm can provide stupid output. In order to prevent the algorithm from “showing you hotels in Ljublana 14 years after you’ve been there,” you need to make sure to give your algorithm a sensible lifetime, Jakobsen says (Interview 2, May 2016).

Programming the News

135

The decision to start using an algorithm in the first place is not always as clear-cut as one might think. Christian Gillinger, who works as head of social media at Sveriges Radio, says of the process of developing the concept for a socially generated top list to be shown on the front-page of the public service broadcaster’s website that “all of a sudden we discovered that what we were doing was creating an algorithm” (Interview 17, May 2016). The social media team had been given the task of finding new ways of promoting socially engaging content. They started down the obvious route, looking at how well certain programs did in terms of likes, shares, and comments. However, they quickly realized that the list would merely reflect and be skewed toward the most popular programs. As Gillinger says, there is nothing particularly social about merely looking at the number of likes and shares for a particular program. If you are a small program, say a Kurdish program with 2000 listeners in total, and you still manage to mobilize 1900 people to share or like you posts, then you are really doing a good job in terms of being social, he adds. Initially, “we didn’t know that what we were developing was an algorithm, but we knew we could not leave it up to the diferent news sections” to define the social; “it had to be done with the help of mathematics.” Because “an algorithm is just a mathematical formula,” the critical task is to decide what the algorithm is supposed to reflect (Interview 17, May 2016, emphasis added). For Gillinger, making decisions about what the algorithm is supposed to reflect is critically dependent on the people involved in actually developing the algorithmic logic. While algorithms are often accused of foreclosing diversity in terms of access to content, the diversity of people designing the algorithms is much less discussed. The more people who think about algorithms and help develop them inside news organizations, the beter, Gillinger argues, suggesting that algorithms require diversity in the teams that create and are responsible for them.

While the algorithm reveals itself as problematic in the initial phases of development and implementation, it may not always solve problems, and it may also create news ones. Sometimes, algorithms simply personalize too much, or they turn out not to automate things as much as anticipated. The news-ranking algorithm at the center of Omni’s editorial infrastructure does not merely support the editors’ workflow; the editors also have to support the algorithm in various ways. As the field of infrastructure studies has taught us, infrastructures require ongoing human commitment and maintenance work (Bowker et al., 2010; Star & Ruhleder,  1996). Infrastructures can be understood as enabling resources; yet, in order to function properly, they are dependent on “human elements, such as work practices, individual habits, and organizational culture” (Plantin et al., 2016: 4). Initially, the idea of assigning the function of curating the front page at Omni to an algorithm was to reduce the need for manual labor. As “Frank,” one of the editors working at Omni, explained to me while I sat next to him observing him at work, the idea is that editors only have to set the news and lifetime value once based on their journalistic know-how; and, then, the algorithm would take care of the rest (Interview 12, March 2015).13 Despite the relative simplicity of the algorithm at work, it turned out that not only did the algorithm not eliminate the need for humans in the workplace, it also, somewhat paradoxically, seemed to put new tasks on the table. As was described earlier, part of the work that editors do at Omni is to assign news and lifetime values to each news story, so that the news-ranking algorithm is able to process it accordingly. Based on these values and a few other signals (i.e., user’s explicit news setings and diferent weights assigned to diferent news sections), the algorithm works to compute the right order on the front page and the right length of time a news story is allowed to remain in that position.

136

I F . . . T H E N

Although the algorithm is programmed to curate the front page, the editors need to check on the algorithm’s performance continuously. As “Emma” (another editor) explained, they need to keep track of how the stories actually appear on the front page and the extent to which the front page reflects the front pages of other major newspapers (Interview 13, March 2015). In other words, editors cannot simply assign their news and lifetime values and trust that the algorithm will do its part of the job correctly. It needs to be checked constantly. It is not just that editors need to check the actual front page to see how their stories are displayed. They also need to make sure that the collective product—Omni’s front page—comes together nicely. In order for the front page to be displayed in the desired way, editors sometimes have to change the initial values given to a news item after it has already been published. This practice, more commonly referred to as “sedating the algorithm,” is by no means uncontroversial among the news professionals at Omni. Having to “correct” the algorithm by manually refurbishing the front page seems to go against the purpose of employing an algorithm in the first place. As Gustafsson and several of the editors told me, the whole idea of assigning values is premised on the notion that the values represent more or less objective and stable journalistic criteria of newsworthiness. For example, changing a story that would, in reality, only merit a news value of 3 into a 5, which normally indicates breaking news of major importance, may not just feel counterproductive. It can also complicate the workflow itself. There is a certain sense in which the initial values cannot be taken as a given, and that you as an editor have to be prepared to change and adjust your own valuation practice to accommodate the practices of your colleagues and whatever story they are working on, in order to achieve the final product that is the front page.

While everyone at Omni seems to agree that “sedating the algorithm” may not be ideal, there is also the sense in which these unintended consequences become an opportunity to engage in an ongoing conversation about what news should be and how the editorial news flow should ideally be organized. At Omni the meaning ascribed to algorithms had less to do with the algorithm in the technical sense and more to do with the algorithm as an infrastructural element. Despite describing the algorithm as “labor-saving device,” for “Frank,” the algorithm seems to be merely one among many other elements that allow him to work in the way that he does— “without anyone telling you what to do” and with a great deal of responsibility and freedom (Interview 12, March 2015). The unintended consequences of algorithms notwithstanding, the same consequences may on a diferent occasion be turned into a welcome argument in favor of using algorithms. The need to check up on the algorithm’s performance as evident from the editorial workflow at Omni becomes at the same time an argument in support of human labor and know-how.

Programming the News

137

Humans Have Heart, Computers Do Not

Algorithms, it seems, have their obvious limits, especially when people position algorithms in opposition to human capacities. Humans have heart and instinct, while computers do not. The ability to judge newsworthiness is seen as a distinctly human quality and a central factor of what journalism is about: craft. It is contingent on journalistic instincts and professional know-how. While some of the interviewees, as we have seen, think there has been too much emphasis on “human instincts” in journalism, “gut feeling” is still an important part of the professional self-understanding of journalists and editors (Schultz, 2007). The journalistic “gut feeling” constitutes a tacit form of knowledge about the newsworthiness of a story. To most practitioners, it is so self-evident that explaining what this gut feeling actually amounts to can be dificult. Even textbook authors charged with articulating the nuances of news practice for newcomers, describe the end-product of journalistic socialization as “instinct” (Coter, 2010: 77). Indeed, many of the interviewees rely on well-known journalistic discourses of professionalism when talking about the role of algorithms in journalism.

As the role of information gatekeeping increasingly lies in the hands of engineers and computer programmers, journalists and news professionals are faced with the dificult task of revisiting professional codes of conduct, and defining what the role of journalism should be. Major concerns in these discussions are whether robots will supersede human labor and what increased automation means for job security. When prompted to think about this, many of the informants talked instead about the limitations of algorithms. While algorithms promise to make the editorial workflow more eficient, this is also where the relationship between algorithms and automation hits its boundary. Algorithms, most interviewees contend, are basically just math, formulas, and calculations. Geir Larsen, then head of analysis at Norway’s most read newspaper $V G ,$ notes how the growing amount of data is just a fact, something news organizations have to deal with sooner rather than later. Yet, the influence of data and algorithms should not be exaggerated. “An algorithm is just a form of programmed logic,” Larsen replies when I ask what kind of algorithms VG is using. “Big data or not,” Larsen continues, “we do a lot of interesting things with small data, too. Whether the data set is big or not is not that important.” At the end of the day, “the most important working tool you have is your journalistic instinct” (Interview 20, March 2014). Even at a news organization such as Omni, which fundamentally relies on algorithmic intermediaries, the computer’s inability to infer contextual information or understand irony means that “there is no chance that algorithms will ever be able to replace the human heart and brain in knowing what is best for the readers” (Interview 8, April 2014). This juxtaposition of human judgment and computational mechanisms does not just reveal a way of thinking about algorithms as mechanical and inferior; it also speaks to the desire of maintaining journalists’ professional environment. As Larsen suggests when I ask about the prospects for automating the curation and presentation of the front page at VG, “We have thought about it, but the answer is no.”

138

I F . . . T H E N

As do many of the other informants, Larsen emphasizes the “democratic” function of news and the journalistic values. While he is open to the possibility of using algorithms to automate some limited and carefully controlled parts of the newspaper, he insists on the fundamental role of human editors in deciding what readers should know. The newspaper’s role, in Larsen’s view, is not to serve content based on individuals’ past clicking behavior. If something important just happened in the world, “it is our role to inform people about it and say ‘this is important, this is what you need to understand’ ” (Interview 20, March 2014). Key to this public information function is maintaining people’s trust in the newspaper as a credible source. The editors and managers talk about their brand name and the importance of having a distinct editorial line that people recognize. For some, the advent of an algorithmic logic in journalism seems to threaten the unique and recognizable editorial line that people have come to know. As Møller says when I ask about the prospect of using algorithms to personalize Politiken’s front page: “We want Politiken to remain a credible media outlet. The best way to accomplish that is through our own prioritizations, not through something automatic.” People come to Politiken because they know what “we stand for,” Møller adds. He explains, “One of our greatest values as a newspaper is that it is created by people with a specific worldview” (Interview 5, March 2014).

The informants’ concerns about the damaging efects of algorithms are familiar ones. Over the past decade, concerns about the algorithmic creation of so-called filter bubbles and echo chambers—the idea that users only get served more of the same content based on their own clicks—have regularly been voiced (Pariser, 2011). These discourses constitute an important backdrop against which the informants express their own concerns. There was a sense in which the linking of journalism and the computational almost required informants to reflect upon the perceived dangers of algorithmic personalization. While Jens Nicolaisen, digital director at

Programming the News

139