You are on page 1of 31

Communication Theory (accepted for publication)

Algorithmic Personalization and the Two-Step Flow of Communication

Oren Soffer

This study examines the relevance of traditional mass communication’s two-step

flow-of-communication theory in relation to algorithmic personalization. I

compare the two-step flow theory’s concept of personalized content through

opinion leaders with the current notion of personalized algorithms, arguing that

opinion leaders and algorithms both function as gatekeeping agents. I also discuss

the nature and role of peer groups in the two cases, arguing that while in the

original theory, groups were seen as relatively solid (family, friends, and work

colleagues), groups in the algorithmic era are much more liquid, transforming

according to data inputs and users’ behavior. Finally, the article also considers

differences in the source of authority of opinion leaders and algorithms in both

eras, as well as the different social settings and public awareness in the second step

of the communication flow.

The current media environment poses new and comprehensive challenges for communication

scholars. Many theoretical assumptions in communication studies—established as this field

evolved and matured throughout the 20th century—have seemingly been undermined.

Fundamental changes in communication technologies in recent decades necessitate a re-

evaluation of common theories developed in the context of "traditional" mass communication

(Chaffee & Metzger, 2001; Napoli, 2010; Soffer, 2010, 2013). My underlying assumption in

this theoretical paper is that "old" media theories—the array of ideas relating to traditional

mass communication—are important tools in the critical study of digital era. How we adapt

1
such theories to preserve their relevance often provides important indicators of the

similarities and differences between the traditional mass communication era and the current

data environment.

I will focus in this study on a seminal theory in communication studies: that is, the

two-step flow of communication, usually attributed to Katz and Lazarsfeld (1955/1966). The

innovation of this theory lies in its highlighting the role of interpersonal relations in the

mediation of mass communication outlets. The theory assumes that most people get their

information not directly from the media but instead through personal sources, thus

emphasizing the role of social groups and opinion leaders in this mediation process.

While attempts to examine the role of the two-step flow theory in the digital media

environment have so far focused mainly on the socio-political role opinion leaders play in

social networks, I will argue that the mediation of opinion leaders is often replaced by

automated, computational two-step flow mediation. This algorithmic computational

calculation imposes grouping clusters onto individuals based on data calculations, taking into

account users’ online history. The role and definition of the group or relevant other peers

have changed in the digital environment. Group clusters are much more instrumental,

temporal, dynamic, and present-oriented than the groups of families, friends, or work

colleagues imagined in the original theory. Yet, groups have not lost their relevance, playing

an important role in algorithmic calculations.

I argue that the attempt to "personalize" the algorithmic flow—that is, giving it an

interpersonal character such that it seems the algorithm knows the users and their tastes—

continues and strengthens trends in traditional mass communication: to give the audience the

feeling of being personally addressed by the mass media (Horton & Whol, 1956; Scannell,

2000). Furthermore, in my discussion of algorithmic personalization in the context of the

2
two-step flow theory, I highlight the crucial role of algorithms in the current socio-political

sphere as well as the risks of putting such an integral part of decision making in the public

sphere into the hands of machines.

The Two-Step Flow Theory

The idea that mass communication content is mediated through personal contacts emerged in

studies of the US presidential election campaign in the 1940s. These studies found that

personal relationships had a stronger effect than the mass media on decision making during

the campaign. Voters revealed that their decisions were often influenced by family members,

friends, and co-workers through conversations held in different social frameworks

(distinguished by age, occupation, or political orientation) (Katz, 1957). These results pointed

to the importance of personal grouping in the mediation of mass communication content. In

each group, certain people played a central role in mediating relevant information between

the mass media and the group (Katz & Lazarsfeld, 1955/1966; Katz, 1957). These leaders,

having relatively large circles of contacts and interest in specific domains (Winter &

Neubaum, 2016), were asked their opinions on issues being discussed in the media, and in

this way influenced others. This mediated influence of mass communication content differed

from previous theoretical perceptions of the direct "stimulus-response" effect of mass

communication (Laughey, 2007, p. 24).

The insights of scholars like Katz and Lazarsfeld led others to the acknowledgment of

the role of social contacts in mediating information that originated in mass communication,

revealing the important communicational role of opinion leaders. These leaders are an

integral part of each social group and stratum, present in its everyday give-and-take, and their

opinions resemble the views of the people they inform and influence (Katz & Lazarsfeld,

3
1955/1966; Katz, 1957). They are considerably more exposed than other group members to

formal mass communication (Katz, 1957). The two-step flow theory assumes that media

contents often "flow from radio and print to opinion leaders and from them to the less active

sections of the population" (Katz & Lazarsfeld, 1955/1966, p. 32). This flow process is

relevant with regard not only to public affairs issues, but also to other fields, like marketing

or fashion. The characteristics of opinion leaders vary from one field to the next, and leaders

and followers of the same social group might exchange roles in different contexts or spheres

of knowledge (Katz, 1957).

According to the two-step theory, mass communication contents percolated through

the opinion leader’s personal (and therefore also often the group’s) agenda and beliefs.

Opinion leaders were therefore active gatekeepers in the process of exposing media content

to the “passive” majority in society (Laughey, 2007). Their authority was seen to be anchored

in their personal characteristics (for example, in fashion, youthfulness played a major role)

and in the knowledge they were thought to hold (through for example their personal contacts

outside the group). They were seen as role models for other group members: "In other words,

it takes two to be a leader—a leader and a follower" (Katz, 1957, p. 74). While opinion

leaders might be more interested than other group members in matters in which they were

“experts,” it is unlikely that their level of interest will differ dramatically (Katz, 1957).

The two-step flow had an important role in directing scholarly attention beyond the

study of mass media effects on cognitive, emotional, and behavioral levels, which occur

following the exposure of a specific individual to a specific medium. The theory pointed out

that any comprehensive evaluation of the impact of mass media requires taking into account

interpersonal discussions about media content that occur over time (Jensen, 2016). The media

impact is therefore not limited to its audience, as the media contents diffuse through

interpersonal relationships to groups that were not exposed to the original message. This

4
undermines the distinction that often stands at the basis of media effect studies—that is

between those who were exposed to a certain message and these who were not. As Vu and

Gehrau (2010) show in their study of agenda diffusion, interpersonal discussion about certain

content from mass media strengthens the issue’s importance among these who were exposed

to the content, affects the issue’s estimated importance among these who were not exposed to

it, and sometimes even motivated them to consume the content that was the object of the

interpersonal discussion.

Along with recognition of its importance, the two-step flow theory also received

severe criticism (Himelboim, Gleave, & Smith, 2009). Weimann’s (1982) critique of the

theory’s assumptions referred to several issues, among them its neglect of the possibility of a

one-step flow from mass communication to certain people, as well as its focus on the vertical

flow from the media to the people while ignoring horizontal flow of opinion-sharing among

peers or non-media sources. As I will discuss later in the article, these lines of criticism are

significant in considering the characteristics of algorithmic personalization in the new media

era.

The Two-Step Flow Theory in the Digital Era

According to Lievrouw (2009), the history of communication as a discipline is often explored

through the distinction between face-to face and technological-mediated communication. She

sees the two-step flow of communication as a main touchstone in the convergence of

interpersonal and mass communication and acknowledges communication as a network

process embodied in a specific context and human negotiation of media content (Lievrouw,

2009). The next important touchstone she identifies in the relationship between face-to-face

and mediated communication is the phenomenon of digital media technologies. According to

5
Lievrouw (2009), the new media "crisis" encouraged a scholarly discourse that "focuses on

communicative action in context of networked relations and systems" (p. 315)

The seminal role of the two-step flow theory in communication studies, along with its

networked, contextual characteristics, drew scholarly attention to the relevance of this theory

in the new media environments—in particular regarding the role of opinion leaders in social

networks such as Facebook and Twitter. As Karlsen (2015) explains, the diffusion of content

in social networks depends on active users, such as opinion leaders. Thus, the logic of

diffusing content from active leaders to more passive individuals resembles the diffusion both

in the two-step flow and in social networks.

In a experiment examining the relevance of opinion leaders' recommendations about

news content, researchers found that "news shared by friends on Facebook is perceived as

more trustworthy than stories received directly from the outlet" (Turcotte, York, Irving,

Scholl, & Pingree, 2015, p. 529). Furthermore, opinion leaders’ recommendations of specific

news articles increased followers' future attention to other news coming from the same news

source. Another study (Karlsen, 2015), examining the diffusion of content from Norwegian

politicians and political parties on Facebook, found that most followers of these politicians

and parties were opinion leaders—being more active in the online sphere and having more

friends and followers than other people. They were also more active in the offline sphere

compared to the larger population.

Along with these empirical studies examining the relevance of the two-step theory in

the digital social network sphere, scholarly attempts were made to theoretically re-

conceptualize the flow of communication in the digital era. One of these was Bennett and

Manheim’s (2006) one-step flow theory. In their work, Bennett and Manheim examined

changes both in social and economic global structures (e.g., as discussed by Giddens [1991])

6
and in technology and the way users interact with it. Their analysis brought them to conclude

that current social-technological changes push towards individual isolation, undermining the

axioms of the two-step-flow theory and paving the way to a one-step flow of communication.

They suggest that group mediation, which stands at the core of the two-step flow, is much

less significant in the new media era than it was in the mass media era. Observing

individuals’ direct interaction with media outlets that are tailored or channeled to match their

characteristics, the authors argue that interaction between people and technology has

changed: "Their use produces an interaction, not among members of peer groups, but between

the technology and the individual audience member" (Bennett & Manheim, 2006, p. 216). In

the decade since Bennett and Manheim’s paper was published, scholarly awareness of the

role of new digital intermediations has increased. The somewhat naïve scholarly perception

of disintermediation in the digital sphere has eroded in part because of empirical evidence of

the continued relevance of intermediators in social media networks (Turcotte et al., 2015;

Karlsen, 2015; Bravo & Del Valle, 2017; Bergström & Belfrage, 2018).

Another attempt to theorize communication flow in the digital era can be seen in

Thorson and Wells’ (2016) “curated flow” approach. Their model takes into account

individual agency and the level of intentionality in people’s exposure to information.

They distinguish among the following: journalistic curation, which is built on the

central, although eroding, role of journalists as information providers; strategic

curation, in which individuals are targeted directly by professional political

communicators; personal curation, in which the choice of information flow is in the

hands of the individual; socially curated flow, in which people choose the social

networks (and therefore the information) they are part of and exposed to; and algorithm

curation, in which specific information is automatically targeted by algorithms.

7
Thorson and Wells (2016)—while noting that strategic curation follows Bennet and

Manheim’s one-step flow model—at the same time identify social curation with the two-step

flow model. They also note that algorithmic curation overlaps with the other four types of

curation. I will focus here on the algorithmic flow, which indeed includes various levels of

inputs—such as strategic, journalistic, and personal. As I will claim, this meta-algorithmic

mediation can be understood and theorized as a two-step flow, in which all the inputs,

including these related to the “calculated peer group,” are used to personalize the contents

provided to the user. Of course, as will be discussed in the following sections, deep

conceptual adaptations are required in the two-step flow theory in order to apply it to the

process of algorithmic personalization.

Algorithms as Social Agency

Algorithms are sets of defined codified steps that are meant to solve certain problems or

fulfill defined tasks (Diakopoulos, 2015). Their computational processes allow them to make

sense out of large sets of data. An algorithm "takes some value, or set of values, as input and

produces some value, or set of values, as output" (Cormen, Leiserson, Rivest, & Stine, 2009,

p. 5). In the last decade, algorithms have played an increasing role in the media sector in

general and in digital journalism in particular. Among other things, algorithms in media-

related corporations function to (1) determine the contents that users are exposed to and (2)

create content, sometime replacing, at least in part, human agents (Napoli, 2014).

I will refer here to the first category of algorithms. These include applications such as

search engines, which are used to prioritize and rank search results; to aggregate and filter

content, such as news stories or posts on social media platforms; and to recommend systems,

which are based on association of similar definitions as entities (Diakopoulos, 2015; Just &

Latzer, 2017; Diakopoulos, 2016). These algorithms influence the exposure of certain

8
perspectives and voices. Thus, search engines, which rank and organize information, make

some of this information prominent while rendering some of it invisible. This hierarchy in the

presentation of content is part of the construction of modern power relations (Lupton, 2015).

While algorithms might be imagined as anchored in natural computational

calculations, they are often created because of capital interest. Among other things, they are

used to improve the efficiency and accuracy of marketing by taking into account users' online

behavior. Algorithms construct reality in a way that reflects the interests of certain groups

and viewpoints (Kitchin, 2017). While some hoped that the Internet would decrease the

power of gatekeepers, algorithms perform data selection and affect how public opinion is

framed (Just & Latzer, 2017). The role of mass media in constructing social reality is well

acknowledged; however, as Just and Latzer (2017) note, the role of algorithms is different:

Algorithmic selection differs from traditional mass-media selection mechanisms,

which operate with a time delay and are mostly targeted at well-defined general

publics and (mass) markets whose characteristics are known from limited data

sources (representative polls, socio-demographic patterns, TV meters in selected

households) compared to big data in the case of algorithmic reality mining and

construction. The database for algorithmic selection applications consists of both

active consumer input (e.g. feedback) and passive data (e.g. location-based,

clickstream, social contacts). (p. 248)

There is growing awareness of algorithms’ function as agents, as they increasingly take an

active part in shaping broad aspects of everyday practices and attract wide social attention

(Kitchin, 2017; Seaver, 2013). Of course, this focus on technological artifacts as agents in

information retrieval and filtering did not start with the operation of algorithms. The actor-

network theory, implemented with regard to prior eras, posits that “the boundaries between

9
the technical and the social, and between human and machine capabilities, are frequently

contested and always negotiable” (Walsham, 1997, p. 477). Thus, the “networked” perception

of pre-algorithmic information systems highlights, in a similar way as discussions of

algorithms, the complexity of interests and logic in content selection. Yet, it seems that the

algorithmic era further strengthens and formulizes this human / non-human melting pot

(Callon, 2004).

According to Napoli (2014), algorithms function as decision makers and should be

considered like human actors or institutions in their effect on social decisions—as they can

affect users’ preferences and behavior through technological means that are embodied in

organizational contexts. The comparison of algorithms to institutions or human actors

emphasizes the multiple influences and forces at play that determine their operation. In fact,

the common metaphor referring to the human brain as a “black box,” emphasizing its

complexity and the non-transparent, abstract flow from inputs to outputs, is also applied to

algorithmic operation (Pasquale, 2015; Kitchin, 2017). The non-transparent and abstract

nature of algorithms starts with their coding, which usually occurs in the private sphere and

therefore is not transparent to the public. The creation of algorithms involves translating a

defined problem or task into operating principles, and then translating this set of principles

into an array of codes. While these processes are often imagined by the coders as natural and

abstract, the operationalization of the problems and tasks, as well of their pragmatic codified

solutions, relates to certain logic, culture, and political context (Kitchin, 2014).

The complexity and non-transparent nature of algorithms' operation also stems from

their operation in systems of other algorithms and actors. They are, according to Kitchen

(2017), “embodied within complex socio-technical assemblages made up of heterogeneous

set of relations including potentially thousands of individuals, data sets, objects, apparatus,

elements, protocols, standards, laws, etc. that frame their development” (p. 20). Furthermore,

10
algorithms that are based on machine learning are constantly evolving through the interaction

and implementation of new data (Kitchin, 2014; Cheney-Lippold, 2011). They have a

constantly changing nature; even their creators cannot entirely track, explain, or predict how

their choice of design will be translated into operation. Given this complexity, it is not

surprising that end user’s often lack understanding of how algorithms operate (Rader & Gray,

2015).

The increasing importance of algorithms in everyday social and political construction,

on the one hand, and the complexity of the multiple factors affecting their operation and

results, on the other, leads to calls for algorithmic transparency—to reveal the people who

control the algorithms, the criteria for algorithms' operation, information about the

characterstics of input data, the rate of error, and the possible sources of bias (Diakopoulos,

2015). We encounter such calls for algorithmic transparency with corporations’ interests and

desire to preserve trade secrets and avoid damaging the assets that contribute to their

competiveness (Diakopoulos, 2015; Diakopoulos & Koliska, 2016), as well as with concern

over overloading users with information they have no interest in (Diakopoulos & Koliska,

2016). As we will see, algorithms are being used as tools to personalize contents to which

users are exposed. In that, algorithms are in some ways fulfilling the role reserved for opinion

leaders in the original two-step flow theory.

Personalization: Between Mass Communication Era and the Algorithmic Turn

The need to personalize media content did not originate in the era of algorithms. In the early

days of the electronic media, the monological nature of mass media was recognized as a

problem that needed to be addressed. As Scannell (2000) notes, the BBC adopted the

principle that radio programming should not be aimed at a “mass” audience, but rather should

11
give listeners the feeling that it is directed at each individual. He refers to this strategy as

“for-anyone-as-someone structures” that reflect a personalized and a conversational mode

allegedly speaking directly to the individual listener. Such modes were adopted in an attempt

to create a personalized and intimate atmosphere in mass media. Thus, Horton and Whol

(1956) point to the attempt in electronic mass media to create a para-social relationship—a

“simulacrum of conversation” between the media “persona” and audience members.

Although they acknowledged that these relationships are "one-sided, nondialectical and

controlled by the performer" (Horton & Whol, 1956, p. 215), they suggested they can be

imagined in interpersonal terms. The illusion of an intimate dialogue between audience and

persona was promoted by the talk style, appealing to listeners as if they were friends, as well

as through the perception of the “subjective camera,” where the lens became the eyes of the

audience. This “mediated quasi-interaction” blurred the monological nature of mass media,

creating for listeners and viewers a kind of friendship with media personas (Thompson, 1995,

p. 90). Both these strategies—what Scannell (2000) coined as “for-anyone-as-someone

structures” and Horton and Whol (1956) as para-social mode—indicate the perceived

disadvantage of mass communication: that it can never be an interpersonal dialogical one.

Despite this, however, mass communication is often evaluated with reference to

characteristics of interpersonal communication (Lievrouw, 2009).

If we consider the two-step flow theory against this background, the second step in

the process is part of personalizing and adapting general monological mass media production

for a relatively small group of individuals. While the first step in this flow imagines the

transmission of the same content to the mass of people, the effect of such universal flow is

assumed to be limited. There is a need for personal mediation of the content, which is

achieved through the interaction between opinion leaders and group members. The opinion

leader, who has the same social background as the group members, can customize the

12
universal message to suit the group members' habitus and perceptions. Such a leader knows

how to re-frame messages: what to emphasize and what to downplay so that the information

will fit group members. If we interpret the two-step flow in this way, using Scannell’s (2000)

idiom, the second step of the flow takes the “for-anyone" content and further strengthens the

frame of the "as-someone" structure. The two-step flow can be seen as a supplementary

social mechanism meant to compensate for the lack of interpersonal conversation in the mass

media era. Generic mass communication outlets are considered as a starting point for

interpersonal discussion, which is held in small groups of relatively socio-culturally

homogenous people. These people conduct full intergroup and interpersonal dialogue

(distinct from the perceived monological one-sided flow of information in the mass media) in

the same way we imagine discussions with our family, friends and colleagues.

In the digital era, however, personalization processes have new ramifications. Digital

media environments allow much easier and more effective tracking of users' behavior and

preferences than was possible in the mass media era (Sundar & Marathe, 2010). Algorithms

and personalization processes are often mentioned in the same breath. A major step in the

personalization of algorithms occurred in 2009 when Google introduced its personalized

search service. As Pasquale (2015) explains,

Personalization lets us hide annoying relatives on our Facebook feeds, list our

favorite microbloggers, and get updates from crucial RRS feeds. It means that

Google News might give pride of place to baseball, music, or left-wing politics

according to the reputation we establish. It means that Google Search orders its

results according to searches we've made before, through clicks collected by the

Google-owned ad network DoubleClick and through activity on other Google-

related accounts. (pp. 78-79)

13
Algorithms, through recommendations and personalization, construct social realities. In this

way, they are similar to traditional mass communication. But algorithms are intended to

bypass or neutralize one of the main features of traditional media—that is, its mass character.

Algorithms personalize exposure to contents (Just & Latzer, 2016). Instead of allowing the

user to choose from a variety of contents, algorithms often curate the contents. Algorithms

respond in real time to evolving trends in consumer behavior, selections, online activities, and

content consumption, such as location and social contacts (Just & Latzer, 2016). In this,

algorithmic operations are meant to resemble interpersonal interactions, in the same way that

someone who knows the person (an opinion leader, for example) would say to them in a

certain moment and a certain context, "Here, this might interest you!" Perhaps the most

prominent expression of algorithms’ function of determining which is the most "meaningful"

content for users, can be seen in the abandonment by some social networks (e.g., Facebook)

of chronologically ordering of posts in favor of algorithmic curation, sorting the contents

according to the expected interest users will have in them, where the time of posting is only

one component among others in the algorithmic calculation (Luckerson, 2015).

As Thurman and Schifferes (2012) show in their study, news organizations

increasingly use implicit or passive personalization techniques. Unlike explicit

personalization, which relies on users’ inputs, implicit techniques infer users' preferences

from the data collected through monitoring their digital activity. Among others, these

organizations are using data collected on social networks to apply collaborative filtering in

order to preserve their site's relevancy. Some suggest that the use of such tools has the

potential to restrict the variety of ideas users are exposed to and to encourage ideological

segregation, creating hegemonic "echo chambers" (Sunstein, 2009). Worries about the often

hidden effect of filtering and personalization on the heterogeneity and bias in public

discourse, along with general trends of transparency meant to recover public trust in

14
journalism (Diakopoulos & Koliska, 2016), have boosted calls for algorithmic transparency

in news production.

Characteristics of the Two-step Flow in the Digital-Algorithmic Sphere

As explained above, the second step in the flow of mass communication, according to the

two-step flow theory, is to personalize the generic mass communication content. In a similar

vein, algorithms often personalize the results that users are exposed to. We can consider, for

example, a cultural product, such as a new book or movie: A review of this product might

appear on television or in a newspaper, which might inflame a discussion in a relevant social

group. The opinion leader, who shares the same cultural habitus as other group members and

is acknowledged by them as a cultural authority, will share his or her opinion about the movie

or book. Using familiarity with the group members and their cultural tastes, the opinion

leader might encourage them to read the book or watch the movie. This kind of group

discussion takes a generic message, which appears on mass media, and customizes it to

individuals that form the group. The interpretation of the review in the mass media might

differ among different groups of the general population: according to age, socio-economical

and cultural background, level of education, and so on.

A similar scenario is still possible today. Mass communication continues to play an

important role in setting contemporary society’s agenda. In the digital era, the second step of

the communication flow can take place in social networks. In such frameworks, opinion

leaders still play an important role. Furthermore, there is no reason to believe that the

exposure to content on social networks would not percolate through interpersonal discussion

to wider circles than those that were initially exposed to the message, as was discussed in the

context of the original two-step flow. Taking this into account, we can return to the example

above of the dissemination of information or recommendation of the book or movie, but here

15
it is provided to the user through algorithmic processing (through active search or content

recommendation). On the surface, this manifests a direct streaming from the computerized

source to the user. As Bennett and Manheim (2006) observe, interpersonal and direct group

mediation of the content is missing in this process. The only players are the user and the

digital device. The speedy processing of algorithms and the intimacy of electronic device

being used strengthen the feeling of one-step direct flow. Moreover, the algorithmic

operation, which stands behind the information curation, is not visible. Algorithms run in the

background. As Introna (2016) notes, they are "organizing our lives and opportunities

without our explicit participation, and seemingly outside of our direct control" (p. 18). In fact,

evidence shows that some users are unaware of the algorithmic presence and its role in

content curation on social networks (Eslami et al., 2015). Yet, it would be wrong to ignore

algorithmic mediation, along with its social and grouping implementations, just because the

process is rapid and invisible. The fraction of a second that the algorithm runs behind the

scene has been described as a step of mediation: a technical step (de Vries, 2010). However,

it seems that this stage is far from having technological implications only.

While the two-step theory has been criticized for not taking into account unmediated

communication flow or different flow patterns such as horizontal flow in the digital era, there

is often no possibility of bypassing the algorithmic processing or changing the flow pattern.

As Napoli (2013) argues, the complexity of the digital fragmented environment results in a

total dependency of users on algorithms in the navigation process. In fact, navigation that is

not algorithmically driven is, in most cases, not possible or effective.

The algorithmic mediated flow of communication usually occurs in the private sphere,

through individual use of a digital device. Yet, this interaction involves other algorithmic

identities. Algorithms construct reality through their choices. The perception that algorithm

results are personalized points to the calculation of a specific user’s character and behavior.

16
Yet, because algorithms predict behaviors in order "to tell the audience what interests them,

without relating to the actual interest of the audience" (Helberger, 2016, p. 193), they rely on

the calculation of digital behavior of relevant other users. Moreover, while the categorization

of algorithmic users is very dynamic, providing mobile "labels” that are maintained as long

they prove to be working (de Vries, 2010, p. 81), these categories, which need to be

translated and sold to advertisers, must correspond with traditional marketing categories.

Thus, as Cheney-Lippold (2011) argues, "online a category like gender is not determined by

one's genitalia or even physical appearance. Nor is it entirely self-selected. Rather, categories

of identity are being inferred upon individuals based on their web use" (p. 165). Gender

composition in algorithmic identity might change with time and data and present math-based

calculations of the two gender categories (Cheney-Lippold, 2011). Yet, in the end, such

algorithmic identities are used for marketing and will likely refer to traditional categories of

distinctions between men and women. In addition, certain algorithmic identities can be

directed to different algorithms. As Seaver (2013) explains,

I recently interviewed a scientist who works for an online radio company. When I

asked him about the playlist algorithm that picks which song to play next, he

corrected me: there is not one playlisting algorithm, but five, and depending on

how a user interacts with the system, her radio station is assigned to one of the five

master algorithms, each of which uses a different logic to select music. (p. 5)

The application of the two-step theory to the algorithmic operation helps us to understand the

changing characteristics of the role and meaning of the relevant peer group in the

communication flow. The group has not disappeared in the algorithms' mediation; but it is

changed. Following Zygmunt Bauman (2000), I argue here that the group in the algorithmic

process is much more liquid in its nature than that that was imagined in the original two-step

theory.

17
Bauman (2000) distinguishes between the modern and the late modern eras. He

describes modern social structures as “solid,” meaning that they maintain their characteristics

over time. However, in late modernity these structures are melting and changing their nature:

The solids whose turn has come to be thrown into the melting pot and which are in

the process of being melted at the present time, the time of fluid modernity, are the

bonds which interlock individual choices in collective projects and actions—the

pattern of communication and coordination between individually conducted life

policies on the one hand and political actions of human collectivities on the other.

(Bauman, 2000, p. 6)

Bauman’s metaphor regarding the liquidity of late modern social structures has been

applied to distinguish between modern structures and their undermining in late modern

culture, for example, in the concept of digital liquid consumption (Bardhi & Eckhardt,

2017), digital liquid language, which is a so-called lighter form of written language

identified by strong oral influences in chats or texting (Soffer, 2013) and to the

algorithmic culture in general (Introna, 2016). In a similar vein, while the groups of

family members, friends, and co-workers imagined in the two step-flow theory are

relatively solid and have clear characteristics in the modern social structure, the group’s

curation in algorithmic operation is much more fluid. The calculated peer group whose

digital behavior is used to personalize the results is constantly changing. Changes occur

due to a number of factors. They can be the result of organizational amendment of

algorithms’ codes. They are also constantly affected by the user’s own behavior, or even

location, due to the transactional nature of communicative act. Such development or

changes in user behavior will result a shift in the algorithmic output and its identification

with a certain category. The same will occur in the case of changes in general social taste

or political orientation.

18
This perception of groups as liquid and constantly changing entities fits well into

existing scholarly discussion on algorithmic grouping and identities. As Cheney-Lippold

(2011) argues, "[I]n cybernetic categorization these groupings are always changeable,

following the user and suggesting new artifacts to be visited, seen, or consumed" (pp. 175-

176). Categories created by algorithms do not stem from any social rationale, but rather from

pragmatic calculations (de Vries, 2010). Moreover, users, who interact with the digital device

as “individuals”, might be unaware of the grouping and categorization process. The group is

not "there" as it was in the two-step flow. It likely cannot even be imagined, because the

common denominator is usually unknown. This is quite different from the visible

simultaneous reading of the newspaper that Benedict Anderson (1983) described as

contributing to national imagination. In the algorithm era, the grouping influence is expressed

in the specific content that users are exposed to and—perhaps even more importantly—

through the absence of other content that seems unsuitable for the algorithmic identity of a

given user (Pariser, 2011). Because the algorithm prioritizes content that fits one's

algorithmic identity, Facebook users, for example, might find themselves being fed political

views that match their own views, while the views of the rival camp become invisible.

It is also important to examine the authority of algorithms from the point of view of

the users. As discussed, in the two-step flow theory, opinion leaders are seen by group

members as having knowledge and expertise in the specific mass media content they are

mediating. The specialty of the opinion leaders, which varies across different fields, gives

them the authority vis-à-vis the group members. The source of algorithmic authority differs

from this, yet it is also crucial in garnering the trust of people in algorithmic curation.

Algorithms are far from being neutral: they classify people and nudge their behavior for

capital reasons (Kitchin, 2017). Yet, because algorithms involve machines and complex

calculations, they are often seen as objective and neutral (Introna, 2016). They enjoy the

19
objective aura of an automated, usually poorly understood, process that is operated without

human intervention (Introna, 2016).

The following table crystalizes the similarities and differences between the original

two-step flow and its application in the contemporary algorithmic environment.

Table 1: Comparison of the the original two-step flow theory and its adaptation in the

algorithm era

Parameters Original two step theory Algorithmic two step


Personalization of Occurs through personal Occurs through matching
contents framing of mass media existing content to an
content by opinion leader individual user through
algorithmic calculations and
association between entities
Social setting Interpersonal relationships Individual use of a digital
device
Sequence and relationship Two individual events occur Two imminent and
between the steps independently, with possible dependent steps occur
delay between exposure to simultaneously
media content and
interpersonal discussion of
them
Agency behind the second Opinion leader, acting as a Algorithm, acting as a
step gatekeeper gatekeeper
Source of agency Expertise in the topic under Mathematical, automated
authority discussion "neutral" procedures
Nature of the peer group Relatively solid, based on Liquid, constantly
in the second step personal acquaintance transforming according to
between its members data inputs and user
behavior

20
Awareness of individuals High, as they take an active Potentially nonexistent, as
to the second step role in interaction the algorithm processing is
hidden

This comparison clarifies the important social and political role of algorithms in the current

communication flow, which had been reserved for human actors in the original two-step flow

structure. Human beings are not immune to manipulation. Yet, they have their own

independent beliefs and opinions. They have the capacity to be aware of fairness issues. They

have consciousness. The problem in putting such a large part of decision-making in

mediatization into the hands of a machine is not only due to the algorithms' bias or

operational effects—for example, in creating a filter bubble or an echo chamber (Pariser,

2011; Sunstein, 2009). Algorithms as instruments can be manipulated: they can be used

against the interests of those who created and own them and benefit those who stand behind

the manipulation. In this way, algorithms can act as a Trojan horse at the heart of the

contemporary communication process.

As Arnoldy (2016) argues, the central role of algorithms as decision makers in trading

markets has created new opportunities for manipulation. He explains that traders have always

tried to manipulate each other; but in the algorithmic era, manipulations that would be easily

traced by human agents are not noticed by the "dumb machines." Moreover, the rise of

algorithmic traders along with new techniques to trick them challenges the definitions of

market manipulations and raises new questions about liability, regulation, legal statutes, and

the role of regulative institutions, as well as the technical possibilities of creating more

protected "smarter" algorithms (Arnoldi, 2016).

21
In the same way that we need to recognize the role of algorithms in trading markets,

we need to take into account their possible manipulation in the market of ideas and its

harmful effect on democracy. Here as well, algorithm manipulation continues a long legacy

of political deception and propaganda. What is new here is the central role played by

machines in making decisions about political discourse. And, similar to the case of marketing

manipulations, questions arise here about liability. This can be seen clearly in the 2016 US

presidential election and the reported Russian attempts to bias it (Shane, 2017). Of course, the

manipulators are the first to be blamed. But what is the responsibility of companies such

Facebook and Twitter whose platforms were manipulated? Can they do a better job

protecting their systems from future manipulations? What about the general socio-political

responsibility of giving algorithms such a central role in the communication process and

decision making? Is a regulative intervention needed or possible? All these questions, which

are already on our daily political agenda, are sharpened when we compare the role of

algorithms to that of opinion leaders in the process of communication.

Discussion

It seems we should put some critical thought into the meaning of “algorithmic

personalization” as a popular and seemingly naturalized turn of phrase in social and scholarly

discourse. Personalization relates to the computational calculation of machines, intended to

create value by providing users with information that matches their algorithmic identity and

preferences (Helberger, 2016). This phrase takes technology, along with the power

relationships behind it, and humanizes it using the terminology of interpersonal relationships.

Marcuse’s (1964) discussion on the functional use of language seems apt here: according to

Marcuse, the use of such language "closes the meaning of the thing, excluding other manners

of functioning" (p. 91). Referring to algorithmic computational calculation as a process of

personalization in a sense personifies it. This continues the trends and strategies of traditional

22
mass communication that aimed to give the transmitted contents an interpersonal flavor—the

“for-anyone-as-someone" strategy (Scannell, 2000) or the “simulacrum of conversation”

between the media “persona” and members of the audience (Horton & Wohl, 1956).

As Lievrouw (2009) points out, the two-step flow theory should be seen as an effort “to

bridge the gap between interpersonal and 'mass' communication”(p. 304). The two-step flow

framed the mass media as being woven within a network of interpersonal group relationships.

While mass communication organizations are the gatekeepers, selecting the information

provided to the public, opinion leaders act as secondary gatekeepers, selecting from the

generic content posted by the media those items that are relevant for their group and

providing interpretations of this content. In a similar spirit, algorithms are seen today to act as

gatekeepers, aggregating the content to which users are exposed (Just & Latzer, 2017). In

fact, as Singer (2014) argues, algorithms often play a secondary gatekeeping role: “In a

traditional media environment, items ignored by an editor were not visible to the public at all;

they did not make it past the journalistic gate. In today’s environment, published items

ignored by users will have made it successfully through that gate but may still fail to reach

more than a handful of readers” (p. 66). The roles of opinion leaders in the traditional media

era and algorithms as secondary gatekeepers point to the social role Web algorithms are

supposed to play. In the spirit of Horton and Wohl’s (1956) theory of the para-social

relationship between the individual viewer and the media persona, algorithms can be

described as acting as para-opinion leaders through the personalization process they enact.

Algorithms—through machine learning and individuals’ interaction with them—are assumed

to "know" each user and curate the relevant information for them.

As I have argued here, at the core of this algorithmic interaction stands a process that

relies on the user’s behavior, liquid social categorizations, geographical location, and other

relevant social information. The liquidity of algorithmic mediation, which in its essence is a

23
process of decision making regarding the social and political public agenda, makes it

relatively vulnerable to manipulation. By examining the interaction between individuals and

their digital devices through the lens of the two-step flow theory, the complexity of the

interaction and its social context are much clearer. Such a view also emphasizes the

continuity between the mass communication era and the digital era. As I argued in this study,

these two eras exhibit significant similarities, specifically between opinion leader's personal

mediation of mass media contents in the former, and the function of the algorithm

mediatization process in the latter—evidenced in the personalization of content that is curated

by algorithms.

References

Anderson, B. (1983). Imagined communities: Reflections on the origin and spread of

nationalism. London: Verso.

Arnoldi, J. (2016). Computer algorithms, market manipulation and the institutionalization of

high frequency trading. Theory, Culture & Society, 33(1), 29-52. doi:

10.1177/0263276414566642.

Bauman, Z. (2000). Liquid modernity. Cambridge: Polity Press.

Bennett, W. L., & Manheim, J. B. (2006). The one-step flow of communication. The

ANNALS of the American Academy of Political and Social Science, 608(1), 213-232.

Bergström, A. & Belfrage, M. J. (2018) News in social media. Digital Journalism, 6(5), 583-

598. doi: 10.1080/21670811.2018.1423625.

24
Bravo, R. S. & Del Valle, M. E. (2017). Opinion leadership in parliamentary Twitter

networks: A matter of layers of interaction? Journal of Information Technology &

Politics, 14(3), 263-276. doi: 10.1080/19331681.2017.1337602

Callon, M. (2004). The role of hybrid communities and socio-technical arrangements in the

participatory design. Journal of the Center for Information Studies, 5(3), 3-10.

Chaffee, S. H., & Metzger, M. J. (2001). The end of mass communication? Mass

Communication and Society, 4(4), 365-379. doi: 10.1207/S15327825MCS0404_3.

Cheney-Lippold, J. (2011). A new algorithmic identity: Soft biopolitics and the modulation

of control. Theory, Culture & Society, 28(6), 164-181. doi:

10.1177/0263276411424420.

Cormen, T. H., Leiserson, C. H., Rivest, R. L., & Stine, C. (2009). Introduction to

algorithms. Cambridge.: MIT Press.

De Vries, K. (2010). Identity, profiling algorithms and a world of ambient intelligence. Ethics

and Information Technology, 12(1), 71-85. doi: 10.1007/s10676-009-9215-9

Diakopoulos, N. (2015). Algorithmic accountability: Journalistic investigation of

computational power structures. Digital Journalism, 3(3), 398-415. doi:

10.1080/21670811.2014.976411.

Diakopoulos, N. (2016). Accountability in algorithmic decision making. Communications of

the ACM, 59(2), 56-62. doi: 10.1145/2844110.

Diakopoulos, N., & Koliska, M. (2016). Algorithmic transparency in the news media. Digital

Journalism, 5(7), 809-828. doi: 10.1080/21670811.2016.1208053

Eslami, M., Rickman, A., Vaccaro, K., Aleyasen, A., Vuong, A., Karahalios, K., ... &

Sandvig, C. (2015, April). I always assumed that I wasn't really that close to [her]:

25
Reasoning about invisible algorithms in news feeds. In B. Begole & J. Kim (Eds.),

Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing

Systems (pp. 153-162). New York: ACM.

Giddens, A. (1991). Modernity and self-identity: Self and society in the late modern age.

Stanford: Stanford University Press.

Helberger, N. (2016). Policy implications from algorithmic profiling and the changing

relationship between newsreaders and the media. Javnost–The Public, 23(2), 188-203.

Himelboim, I., Gleave, E., & Smith, M. (2009). Discussion catalysts in online political

discussions: Content importers and conversation starters. Journal of Computer

Mediated Communication, 14(4), 771-789. doi: 10.1111/j.1083-6101.2009.01470.x

Horton, D., & Wohl, R. (1956). Mass communication and para-social interaction:

Observations on intimacy at a distance. Psychiatry, 19(3), 215-229. doi: 10.1007/978-

3-658-09923-7_7 .

Introna, L. D. (2016). Algorithms, governance, and governmentality: On governing academic

writing. Science, Technology, & Human Values, 41(1), 17-49. doi:

0.1177/0162243915587360.

Jensen, K. B. (2016). Two-step and multistep flows of communication. The International

Encyclopedia of Communication Theory and Philosophy, 1-11. doi:

0.1002/9781118766804.wbiect186

Just, N., & Latzer, M. (2017). Governance by algorithms: reality construction by algorithmic

selection on the Internet. Media, Culture & Society, 39(2), 238–258. doi:

10.1177/0163443716643157

26
Karlsen, R. (2015). Followers are opinion leaders: The role of people in the flow of political

communication on and beyond social networking sites. European Journal of

Communication, 30(3), 301-318. doi: 0.1177/0267323115577305.

Katz, E. (1957). The two-step flow of communication: An up-to-date report on an hypothesis.

Public Opinion Quarterly, 21(1), 61-78. doi: 10.1086/266687.

Katz, E., & Lazarsfeld, P. F. (1966). Personal influence: The part played by people in the

flow of mass communications. New Brunswick (US) and London: Transaction

Publishers (Original work published 1955).

Kitchin, R. (2014). The data revolution: Big data, open data, data infrastructures and their

consequences. London: Sage.

Kitchin, R. (2017). Thinking critically about and researching algorithms. Information,

Communication & Society, 20(1), 14-29.doi: 10.1080/1369118X.2016.1154087 .

Laughey, D. (2007). Key themes in media theory. Berkshire, England: McGraw-Hill

Education.

Lievrouw, L. A. (2009). New media, mediation, and communication study 1. Information,

Communication & Society, 12(3), 303-325. doi: 10.1080/13691180802660651.

Luckerson, V. (2015, July 5). Here's why Facebook won’t put your news feed in

chronological order. Time. Retrieved from http://time.com/3951337/facebook-

chronological-order/

Lupton, D. (2015). Digital sociology. Abingdon, Oxon: Routledge.

Marcuse, H. (1964). One-dimensional man: Studies in the ideology of advanced industrial

society. London and New York: Routledge.

27
Napoli, P. M. (2010). Revisiting “mass communication” and the “work” of the audience in

the new media environment. Media, Culture & Society, 32(3), 505-516. doi:

0.1177/0163443710361658.

Napoli, P. M. (2013). The algorithm as institution: Toward a theoretical framework for

automated media production and consumption. Fordham University Schools of

Business Research Paper. Retrieved from SSRN: https://ssrn.com/abstract=2260923

Napoli, P. M. (2014). Automated media: An institutional theory perspective on algorithmic

media production and consumption. Communication Theory, 24(3), 340-360. doi:

0.1111/comt.12039.

Pariser, E. (2011). The filter bubble: What the Internet is hiding from you. New York:

Penguin.

Pasquale, F. (2015). The black box society: The secret algorithms that control money and

information. Harvard: Harvard University Press.

Rader, E., & Gray, R. (2015, April). Understanding user beliefs about algorithmic curation in

the Facebook news feed. In B. Begole & J. Kim (Eds.), Proceedings of the 33rd

Annual ACM Conference on Human Factors in Computing Systems (pp. 173-182).

New York: ACM.

Saurwein, F., Just, N., & Latzer, M. (2015). Governance of algorithms: Options and

limitations. info, 17(6), 35-49. doi: 10.1108/info-05-2015-0025.

Scannell, P. (2000). For-anyone-as-someone structures. Media, Culture & Society, 22(1), 5-

24. doi: 10.1177/016344300022001001.

Seaver, N. (2013). Knowing algorithms. Media in Transition, 8, 1-12.

28
Shane, S. (2017, September 7). The fake Americans Russia created to influence the election.
New York Times. Retrieved from
https://www.nytimes.com/2017/09/07/us/politics/russia-facebook-twitter-
election.html.

Singer, J. B. (2014). User-generated visibility: Secondary gatekeeping in a shared media

space. New Media & Society, 16(1), 55-73. doi: 10.1177/1461444813477833.

Soffer, O. (2010). “Silent orality”: Toward a conceptualization of the digital oral features in

CMC and SMS texts. Communication Theory, 20(4), 387-404. doi: 10.1111/j.1468-

2885.2010.01368.x.

Soffer, O. (2012). Liquid language? On the personalization of discourse in the digital

era. New media & society, 14(7), 1092-1110. doi: 10.1177/1461444812439550.

Soffer, O. (2013). The internet and national solidarity: A theoretical analysis. Communication

Theory, 23(1), 48-66. doi: 10.1111/comt.12001.

Sundar, S. S., & Marathe, S. S. (2010). Personalization versus customization: The importance

of agency, privacy, and power usage. Human Communication Research, 36(3), 298-

322. doi: 10.1111/j.1468-2958.2010.01377.x.

Sunstein, C. R. (2009). Republic. com 2.0. Princeton: Princeton University Press.

Thompson, J. B. (1995). The media and modernity: A social theory of the media. Stanford:

Stanford University Press.

Thorson, K., & Wells, C. (2016). Curated flows: A framework for mapping media exposure

in the digital age. Communication Theory, 26(3), 309-328. doi: 10.1111/j.1468-

2958.2010.01377.x.

29
Thurman, N., & Schifferes, S. (2012). The future of personalization at news websites:

Lessons from a longitudinal study. Journalism Studies, 13(5-6), 775-790. doi:

10.1080/1461670X.2012.664341.

Turcotte, J., York, C., Irving, J., Scholl, R. M., & Pingree, R. J. (2015). News

recommendations from social media opinion leaders: Effects on media trust and

information seeking. Journal of Computer‐ Mediated Communication, 20(5), 520-

535. doi: 10.1111/jcc4.12127.

Vaccari, C., & Valeriani, A. (2013). Follow the leader! Direct and indirect flows of political

communication during the 2013 general election campaign. New Media & Society,

doi: 1461444813511038.

Vu, H. N. N., & Gehrau, V. (2010). Agenda diffusion: An integrated model of agenda setting

and interpersonal communication. Journalism & Mass Communication

Quarterly, 87(1), 100-116. doi: 10.1177/107769901008700106.

Walsham, G. (1997) Actor-network theory and IS research: Current status and future

prospects. In A.S. Lee, J. Liebenau, & J. I. DeGross (Eds.), Information systems and

qualitative research (pp. 466-480). IFIP — The International Federation for

Information Processing. Boston: Springer.

Weimann, G. (1982). On the importance of marginality: One more step into the two-step flow

of communication. American Sociological Review, 47(6), 764-773. doi:

10.2307/2095212.

Winter, S., & Neubaum, G. (2016). Examining characteristics of opinion leaders in social

media: A motivational approach. Social Media + Society, 2, 1-12.

doi:10.1177/2056305116665858.

30
31

You might also like