Artificial Intelligence in Digitalized News Media

(1)

Artificial Intelligence in Digitalized News Media

A Case Study on Implementing Recommender Systems to Retain High Priority Consumers at Ekstra Bladet

MSc in Business Administration and E-Business

Master Thesis · August 2dn 2021

Character: 123138 Pages: 52 Supervisor:

Julie Gerlings

Author

Hannah Skovlund Carden, 130393

(2)

Acknowledgments

I would like to direct a special thanks to my thesis supervisor Julie Gerlings for providing feed- back, continuous guidance, and enthusiastic encouragement throughout the entire process.

In addition, a special thanks to the case company and Head of PIN, Kasper Lindskow, for providing support and valuable knowledge for the case study and involving me in the existing process of the research project of Platform Intelligence in News. Lastly, an appreciation to the respondents for their time and valuable insights.

(3)

Abstract

Digitalization has influenced the consumption of news media and the democratic corporatist model. Simultaneously recommender systems are becoming an inherent part of consumers' everyday life as well as the concern for ‘filter bubbles’ being created. On the other hand, recommender systems are providing useful to the consumer helping them navigate through the information overload of today’s society. Ekstra Bladet is currently working on developing and implementing such recommender systems as a part of the Platform Intelligence in News (PIN) project. To investigate how the consumers of the Danish news media are perceiving this technology, this study poses the research question: How can Ekstra Bladet implement recommender systems to retain high priority consumers?

Adapting an interpretive research philosophy and an inductive approach to collecting empirical data enables an understanding of how the phenomena of interest are perceived among stakeholders and consumers. As a research strategy, a single case study was applied by performing qualitative interviews with experts and high priority consumers of Ekstra Bladet. Furthermore, the data collected were analyzed by consolidating codes and concepts into aggregated dimensions. The findings contribute to the field of perception of recommender systems in Danish news media in relation to the PIN Project where a limited research has been done.

In conclusion, high priority consumers are positive towards the implementation of recommender systems at Ekstra Bladet’s digital platform. However, the consumer expresses concern about ‘filter bubbles’ and data collection, even though personal data is needed for the recommender systems to personalize content to the consumers. There is no correlation between which consumer segment the respondents are in, but rather a tendency in relation to gender as the female respondents are concerned with their data being collected and the male respondents are indifferent. Among both consumer segments there exists a paradox between wanting personalized content recommended and the unwillingness of allowing more than ‘only necessary cookies’. Furthermore, it can be concluded that the consumer segments are different, hence their consumption of news is different. Thus, this should serve as an underlying premise for having different approaches when implementing recommender systems at Ekstra Bladet and recommending content to different consumer segments.

(4)

1.Introduction

“The most important general-purpose technology of our era is artificial intelligence, particularly machine learning (ML) — that is, the machine’s ability to keep improving its performance without humans having to explain exactly how to accomplish all the tasks it’s given.” (Brynjolfsson &

McAfee, 2017).

Machine learning algorithms using artificial intelligence can make recommendations and thereby personalize content for individual users, also known as recommender systems (Ricci et al. 2015). The recommendations or suggestions related to the process of decision-making for a consumer are in relation to various actions such as what series to watch or what news to read.

These suggestions are customized to provide useful recommendations on a specific item to help the users navigate in a world with content overload (Ricci et al.2015) For businesses such as Netflix, Facebook, Spotify, and Google, recommender systems have been an essential part of their business (ibid.). Algorithmic recommendations in entertainment media like Netflix influences up to 80% of the consumers when selecting a video and have generated one billion dol- lars in profit due to user retention (Thurman et al. 2019).

As recommender systems are increasingly becoming a bigger part of our everyday life, so is the increasing need for personal data used for tailoring content. As a result, thereof the regulation of data privacy in the European Union GDPR came into play in 2018 to provide the consumer with more rights in relation to their personal data (Ignatidou, 2019). Concerns about privacy and ethics were caused by several data leakage cases and Facebook has been criticized for creating ‘filter bubbles’ where people only get shown content aligned with their own worldview.

(Duke, 2018). These systems have not had the same effect on the Danish news media industry yet. This can be due to the barriers in languages and the speed of changes in the flow of news (Copenhagen Business School, 2020). As a result of this, there is currently limited knowledge within the field of recommender systems in the Danish new media and the causes and effects of implementing such systems.

(7)

1.2 Research Question

Therefore, this study aims to explore how recommender systems can be implemented in news media and benefit both the company behind it and the user consuming it. To investigate this, this thesis is written from a case study approach to contribute with insights to the research project Platform Intelligence in News (PIN), which is currently investigating how the use of artificial intelligence and recommender systems can provide better news experience for news media consumers. The case study is guided by the following research question:

How can Ekstra Bladet implement recommender systems to retain high priority consumers?

Based on the research question this study, therefore, investigates Ekstra Bladet’s high priority consumers; The News Junkie Segment’s and The Larger Consumer Segment’s perceptions and meanings of the implementation of recommender systems. Hence what opinions they have about their data being collected. From these findings, the study provides recommendations for Ekstra Bladet when implementing recommender systems on their online platform.

Based on the research question, this thesis explores news media and the digitalization hereof.

This thesis sheds light on the users’ experiences within the digitalization of news media. This will serve as a foundation for further investigating what risks and opportunities the users perceive in relation to implementation of recommender systems in Danish news media. From these findings, the users’ perspectives can be articulated in a way that contributes to the current understanding of news recommender systems. Furthermore, as this thesis is written from a case study approach it therefore contributes these perspectives to the Platform Intelligence in News (PIN) project in which recommender systems in Danish news media are currently researched and developed.

1.3 Structure of the Project

 In order to provide an overview of the thesis the following model provides a brief structure of the study:

(8)

Figure 1: Structure of the thesis

Pursuing to answer the identified research question, the explanation of the overall thesis structure is provided. This structure aims to clarify the connection between the different sections and their content. Thus the thesis is structured as follows:

Present chapter introduces the motivation of the thesis and the posed research question. Chap- ter 2 presents the literature review on existing literature in the field of news media, artificial intelligence including recommender systems, and the connected ethical questions.

Chapter 3 clarifies the methodology and methods applied for this thesis and explains the con- duction of data, how it has been analyzed, accompanied by the case setting.

1 Introduction

2 Literature Review

3 Research Methodology and Methods

4 Findings

5 Discussion

6 Conclusion

(9)

Chapter 4 presents the findings from the case study followed by a discussion in chapter 5 of the implications of the findings that leads to general recommendations for Ekstra Bladet.

Finally, chapter 6 assembles the conclusion of the thesis and future research.

(10)

2. Literature Review

This chapter is devoted to reviewing a selected part of relevant and existing studies on literature about Artificial Intelligence (AI), News Media, and Recommender Systems. To gain an overview of the existing knowledge related to these concepts the literature search was conducted using search terms such as artificial intelligence, recommender systems, recommender systems AND news media, digital journalism, and automated news. The search was conducted within different databases such as Libsearch, SAGE, Oxford Research Encyclopedias, and Google Scholar. The search included articles where the terms appeared in titles, keywords, abstracts, and full text. The articles were screened to determine their relevance based on their abstracts and full text. In addition to this, the reference lists from relevant studies were used to detect and explore supplementary literature. The following sections will present the most relevant studies located throughout the literature search. First, an examination of the concept and term AI is described as a brief introduction for further review within this field. Second, the context of news media in relation to digitalization. Lastly, the concept of recommender systems is ex- amined in which ethical considerations in relation to AI are emanated from. The reviewed literature serves as a scientific foundation for the thesis. In the following section, the relevant literature found in the review is presented.

2.1 Artificial Intelligence

The term “Artificial Intelligence” (AI) has been discussed from various perspectives over the last couple of years but goes back to 1956 where it was coined at Dartmouth College by McCar- thy (McCarthy, 1960). His definition of the term “... a program has common sense if it automatically deduces for itself a sufficiently wide class of immediate consequences of anything it is told and what it already knows” (McCarthy, 1960, p. 3) is still useful for the understanding of AI today even though no unified definition of AI exits globally (Tzafestas, 2015). The definition emphasizes that a program can make choices that are rational if it is told what it needs to learn.

Building on the perspective that machines are able to learn from data and be so-called “intelli- gent agents” Russell & Norvig’s defines AI as: “[The automation of] activities that we associate with human thinking, activities such as decision-making, problem-solving, learning [...]” and the machines’ ability to act rationally.” (Russell & Norvig, 2016, p. 4). Both definitions of AI

(11)

contribute to the understanding of the phenomenon as a machine that can imitate human thinking when it knows what to look for and thereby help solve problems or automate tasks and learn from experience. This is the essence of what the discipline of AI was founded on - that human intelligence could be defined and articulated so precisely that a machine would be able to learn to replicate it over time (Pradeep, A. K ; Appel, Andrew ; Sthanunathan, Stan, 2019).

2.1.2 AI as an Umbrella Term

The term AI is often used in a context with a description of one technology, which is not in accordance with the description from McCarthy (1960) where machines mimic human intelligence to solve tasks, as well as the scholar Burgess (2018) who emphasizes in his work that AI is a wide concept that includes different technologies and not ‘just one thing’. The term includes ruled-based systems, machine learning, speech recognition, and natural language processing (NLP) (Eggers et al., 2017) and can thereby be seen as an umbrella term for different technologies (Pradeep et al., 2019). According to Brynjolfsson and McAfee (2017), the most important of these artificial intelligence technologies of our era are particularly machine learning.

(Brynjolfsson and McAfee, 2017. This research stream of AI is strongly combined with deep learning and therefore not seen as an individual concept, but rather concepts under the same umbrella as explored in the next section. The purpose of this thesis is not to cover all of the mentioned technologies, but instead to focus on the literature under this umbrella term that is underlying concepts of the two main research streams machine learning and deep learning, to understand recommender systems and how to operate with machine learning.

2.1.3 AI Technologies

Machine learning is the process “by which a computer continually adjusts its output based on its own user experience” (Pradeep, A. K ; Appel, Andrew ; Sthanunathan, Stan, 2019, p. 24).

Hence machine learning uses statistics by error and trial to make ‘self-learning’ algorithms that recognize patterns within a huge batch of data to find similar patterns and make a better output forward the more information it gets (Ibid.). This is done when an algorithm uses data to perform a task and gather responses – either positive or negative – to learn how to solve tasks in a

(12)

similar manner in the future (Sterne, 2017). Hence ML constantly adjusts its actions to improve a given task, while AI sets the initial set of rules to maximize the performance of a task. Accord- ing to Sterne (2017) machine learning is designed to learn and needs human interaction as quality control for biases that the machine does not notice yet.

Deep learning refers to machine learning of deeper multilayered neural networks to perform a modeling task based on big data. This digital network mimics the brain neurons in a biological nervous system that enables the brain to perform all kinds of tasks (Pradeep et al., 2019). This enables a more advanced understanding and learning process for the linear levels of representation that the deep learning models consist of as shown in figure X (Sterne, 2017). Deep learning has led to progress within both image recognition and text- and voice understanding, due to its ability to ‘modeling the mind’ and use the human way of thinking and solving specific problems (Pradeep et al., 2019).

2.2 News Media and Digitalization

Communication has become increasingly digital in today’s world. The same applies to accessing entertainment and news which are being more personalized (Friedrichsen & Kamalipour, 2017). This hasty change of digital media technologies has broad different challenges for journalism in general and especially for the printed news (Hjarvard & Kammer, 2015). During the 20th century, the political press was transformed by having a more impartial position as an

‘omnibus press’ – even though opinion journalism still has elements of political parallelism (ibid.) According to Hjardavard & Krammer (2015) the Nordic countries including Denmark operates from the democratic corporatist model where a strong legal authority has provided “a safeguard against direct political influence on news media, allowing public intervention to go hand-in-hand with an ‘arm’s length principle’ of no-interference in editorial matters.” This principle refers to the legacy media’s function as the ‘Fourth Estate’ hence the press complements the three official powers of the state: the legislature, the judiciary, and the executive (Svalgaard, 2017). Ignatidou (2019) stresses the importance of the media’s role when implementing new technologies, thus the press plays a crucial role in society as it shapes the public and the political sphere.

(13)

Both digitalization and commercialization have affected news media by putting on pressure on their revenue streams. Hjardavard & Krammer (2015) points out that the advertising revenue stream has been taken over by businesses like Google and the lack of willingness to pay for news, as they can be found somewhere else for free has influenced the subscription revenue stream. This results for the public news media in the difficulties of on one hand running a private business and on the other hand serving the public interest while at the same time competing with commercial newcomers. According to Søndergaard & Helles (2014), the legacy news- papers have joined forces with other commercial media and are criticizing public service media for unfair competition, hence they are providing unpaid news free of charge. Therefore, the public and private media in the Danish field of media are competing against the internal actors within the field and not only against the newcomers. On behalf of this, the democratic corporatist media model has become pressured, and the business of news is changing (Hjardavard &

Krammer, 2015) as Hallin & Mancini (2014) argues:

Commercialization [...] is clearly shifting European media systems away from the world of politics and towards the world of commerce. This changes the social function of journalism, as the journalist’s main objective is no longer to disseminate ideas and create social consensus around them (Hallin & Mancini, 2014.)

The rise of digital media has pushed the news media toward a more liberalist- and commercial model. This change in the media scene has, according to Ottosen & Krumsvik (2009) made Nor- wegian journalists experience a bigger pressure on delivering news articles that generated revenue, hence got more clicks and thereby a reduction in the quality of news.

2.2.1 Personalization in News Media

Due to the rapid technological innovation and the digitalization of news, more content is being personalized and recommended to online users. Using algorithms in interaction with custom- ers is nothing new. Algorithms have played an active role within journalism since 1970 (Zamith, 2019) and by 2010 semi- or fully automated forms of sharing, filtering and gathering news have become an important part of the news media field (Thurman et al., 2019). These automated processes can be used to test news media content and learn what the user requests more of and

(14)

useful insight in their users’ behavior, personal characteristics, and social networks. Hence, these algorithms can make recommendations based on user information (ibid.).

Through automated journalism hundreds of thousands of articles are automated each year (Za- mith, 2019) and news media businesses are now experimenting and extending the personalized news provided (Thurman et al., 2019). According to Zamith (2019) there is “a trend toward algorithmically enabled personalization.” (Zamith, 2019, p.1) which he pinpoints is making a shift in focus from the shared importance in journalism to a more individualized news experience. A proof of concept is the case with the New York Times and their ambitious project that customized the delivery of news (Thurman et al., 2019).

As algorithmic recommendations become more predominant, consumers’ perceptions of these are of interest. There are different opinions related to algorithmic news recommender systems.

Thurman et al. (2019) stress that there is significant variation in the beliefs which are connected to factors such as mobile news access, trust in news, paying for news, age and concerns about privacy. The scholars also found that the users rather have the algorithm selecting articles based on their past interactions than an editor. Furthermore, overall gratitude towards algorithms among the participants was found and could differ in the context of the data that drives the algorithm. (ibid.). In addition, Bodó et al. (2019) contribute from their study with other factors that affect the value of personalization which include the diversity and depth of recommendations and what they refer to as “concerns about a shared news sphere” (Bodó et al., 2019, p.1). However, these factors are bound to peoples’ level of education. Less educated people show less concern about diversity in their news recommendations. According to Bodó et al. (2019), people do not want personalized content at all costs (ibid.). Awad and Krishnan (2006) contribute with the perception of benefit and outcome are strongly connected to the consumers’ willingness to partake in online personalization. They argue that useful personalization can motivate the consumer to give personal information even though they might have privacy concerns.

(15)

2.2.2 Automatization and Democracy

As stated in the previous section, algorithms that personalize digital content are omnipresent (Bodó et al., 2019). These recommender systems can be useful and necessary (Gauch, 2007).

In relation to enabling user autonomy and information overload, other scholars, nevertheless, point at threats in relation to personalization (Borgesius et al. 2016). Bodó et al. (2019) men- tion the concern about echo chambers:

“The current media law and policy debate (Borgesius et al. 2016; Yeung 2017) about echo chambers (Sunstein 2002) or “Daily Me’s” (Negroponte 1995) is concerned that personalized recommendations are opaque in how they filter/recommend information to individual users, and that this filtering may lead to a number of undesirable consequences.” (Bodó et al. 2019, p.1)

Possible consequences could be the reduction in diversity and the development of ‘filter bubbles’ where news consumers are only shown content that reaffirms them in their own standpoint and worldview (ibid.). These consequences would conflict with information society as the access to diverse information and the participation in political decision making based on – what Habermas (1989) refers to – as the shared public sphere would be threatened (ibid.). Echo chambers and filter bubbles are short metaphors for the public fear that information users con- sume online may be limited (Kitchens, 2020). In accordance with Helberger (2019), recommender systems can be a threat for news media due to the influence they can have on democracy and polarization, but also offer opportunities to the media's democratic role. This depends on the use of the algorithm and the importance of implementing it with an understanding of the democratic values it can bring, and not solely focusing on commercial purposes (Helberger, 2019). She stresses the importance of context when making recommender systems integrate democratic values and the relation between context and values at stake as these cannot be stud- ied in isolation (ibid.).

In accordance with Ignatidou (2019), the implementation of AI technologies needs to be exam- ined carefully due to the responsibility of the legacy media as the ‘Fourth Estate’. This should be done in line with the values of the news media. Furthermore, she states a lack of regulation according to the use of AI, like recommender systems, used by news media. Zamith (2019) sub-

(16)

in news media (ibid.). Both scholars emphasize the lack of transparency within the field which makes it difficult to endorse the accountability of algorithms as they can spread misinformation (ibid.) and influence democratic elections (Ignatidou, 2019). According to Ignatidou (2019) codes of ethics will not be enough – there need to be regulatory frameworks put in place world- wide. Hence the legislation in Europe with GDPR and restring of personalized commercials are not efficient for personalized journalism (ibid.).

2.3 Recommender Systems in News

This section explores recommender systems in connection to news media. This selection of literature is deemed important to get an insight into the development and implementation of recommender systems and the role they play in the field of news. To get an understanding of the news consumers' perception of these systems, it is essential to unfold their design and their influence on online users. Hence the aim of this study is to research how the consumers of Eks- tra Bladet interpret the implementation of recommender systems. Therefore, the literature within this section has been collected in relation to scholars who explore the influence on users that recommender systems can have. First, a definition of the term ‘recommender systems’ is clarified. Secondly, the development and the purpose of recommender systems are unfolded.

Subsequently, recommender systems are presented from a consumer-oriented perspective.

2.3.1 Shift in Paradigm

According to the definition from Ricci, Rokach and Shapira (2015), recommender systems “are software tools and techniques that provide suggestions for items that are most likely of interest to a particular user.” (Ricci et al., 2015, p.1). These suggestions can be related to different kinds of decision making such as what news to read online. Recommender systems normally focus on a specific type of item (e.g news) and the recommendations are customized to provide valuable suggestions for that specific item (ibid.) Another similar definition of recommender systems is found from Sterne (2017) who defines the term as a subclass of information filtering system aiming to predict what preferences a user has in terms of rating a given item based on sophis- ticated machine learning algorithms. Both scholars acknowledge the importance of the user's perception of recommendations.

(17)

According to Seaver (2019), the paradigm of recommender systems has shifted over time. Rec- ommender systems at the early stage were aimed to precisely create predictions. These recommendation systems became difficult to improve as they varied a lot and the uncertainty made them unstable. To cope with this the researcher made use of more metrics within the data to measure user satisfaction. Seaver (2019) refers to this as the ‘paradigm of captivating metrics’

that we are currently in where the target is the consumer. In order to captivate a user “one needs to persuade the target in a way that suits their psychological behavior patterns” (Seaver, 2019. p.4). Seaver (2019) makes an analogy to the way a hunter learns to capture a target by making a trap. The same applies for developing recommender systems where human behavior patterns are essential to understand to captivate or hook the user. According to Seaver (2019) recommender systems can thereby be seen as traps which Rakova and Chowdhury (2019) explores in their research. They derive a metric called ‘barrier-to-exit’ which is the representation of “the amount of effort a user needs to expend in order for the system to recognize their change in preference.” (ibid, p.1). This metric cannot be generalized as it differs between users and is meant to be used on an individual user level. Even though these ‘barriers-to-exit’ are individual, they seem to increase as the amount of information offered to users makes the process of decision making relying more on recommender systems (Ricci et al., 2015). According to Rakova and Chowdhury (2019) using these captivating metrics can increase the user retention and getting the user ‘hooked’ on continuous detailed recommendations based on their data.

2.3.2 Recommender Systems and Consumers

According to the scholars Bucher & Helmond (2018) the perception of recommender systems from a consumer point of view is thoroughly connected to what the technology allows the consumers to do. van Dalen (2020) contributes with his study on the perception of the recommender system in news from a Danish user perspective. He found that users were more positive towards collaborative filtering than solely algorithmic recommendations (van Dalen, 2020). Findings from the music industry and the scholars Goldschmitt and Seaver (2019) are aligned as the users were more positive towards a recommended playlist made by humans than solely algorithms. These studies indicate that recommendation from algorithms with human interaction creates more value for the consumer. In addition, Fast and Horvitz (2017) explored in their study “Long-Term Trends in the Public Perception of Artificial Intelligence” that the

(18)

views on AI had been more optimistic, however Fast and Horvitz (2017) emphasize the negative impacts of AI including ethical concerns and the loss of control in regards to privacy also had increased in the recent years of the 30 year period of the study.

Another perspective on consumer perception of recommender systems is presented by Shklov- ski et al. (2014). Their findings refer to the ‘privacy paradox’ as users are perceiving the necessary data collection for recommendations as‘ creepy’ all though at the same time have an interest in personalized content (ibid.). According to Shklovski et al. (2014) the data collection behind the recommender systems is an ‘abstract concept’ for the users. This privacy paradox and the new paradigm of recommender systems as captivating metrics are facing privacy concerns from a consumer perspective which are some of the negative aspects of recommender systems (Goldschmitt & Seaver, 2019).

2.3.3 Ethics in The Field of Artificial Intelligence

As stated in the previous section the benefits of personalized recommendations are multiple, both for consumers and providers. Nevertheless, many scholars including Ricci et al. (2015) also emphasize in the previous part of the research ethical concerns about risk and the loss of user privacy regarding recommender systems and AI. The risk of privacy breach refers to the need of collecting and storing personal user data when using recommenders (Ricci et al. 2015.).

To provide personalized content the recommender needs data for generating (quality) recommendations to the user (ibid.). This loss of privacy Goldschmitt and Seaver (2019) refer to as the concept of ‘Little Brother’ where humans are recorded through observation. This could be when watching a movie or playing a song (ibid.). This concept could indicate that consumers may have their own opinion about the sensitivity of their personal data and the action they are taking to protect their data privacy (Ricci et al., 2015). Hence consumers' perception of privacy and ethical facets is important to understand.

Applying rules or ethical standards within AI can be difficult as the ethical stance in society has changed (Øhrstrøm, 2014). Powers and Ganascia (2020) state five different challenges in connection to ethics in AI including which ethical approach to take or how to implement machine ethics. The findings in their research reveal the need for procedures to incorporate ethics

(19)

for specific guidelines as a ‘universal set of rules’ to endorse an ethical standard for AI technologies (ibid.). In contrast, Hagendorff (2020) indicates that there already are guidelines for ethics in AI, but in practice, these do not work. Hence, the lack of interest in using these guidelines is bound to the missing consequences (ibid.). Tang & Winoto (2015) reiterate that in addition to guidelines the collaboration between machines and humans is a crucial part of providing an ethical framework. Within ethics in AI Jobin et al. (201) also emphasize the importance of transparency and the need for guidelines hereof that indicate implementation strategies of AI technologies. According to Stohl, Stohl, and Leonardi (2016) there is a ’transparency paradox’ between viewability and transparency that indicates less transparency as too much information is being visible in contrast to the notion of information being transparent when it is easily accessible. The different viewpoint indicates an unclear perception of ethical standards within the field of AI and how they should be put into place. As Hagendorff (2020) emphasizes the practical implications for ethical guidelines and Jobin et al. (2019) perceive the need for guidelines, the literature indicates that there are indifferent perceptions of how ethics and AI should be connected.

(20)

3. Research Methodology and Methods

This section presents and elaborates on the methodological foundation for the project and ac- counts for the methodological choices made. First, a brief introduction and explanation of the research philosophy which includes a definition of the selected methodology. Secondly, the rea- soning for the selected research approach and a detailed elaboration on the selected methods.

Thirdly the research strategy is presented by a single case study data. Finally, an explanation of the collected data and how it has been gathered and interpreted including the integrity of the provided findings and the selected data analysis will be presented.

3.1 Philosophy of Science

This study adopts an interpretivist stance. The philosophy of science serves as the underlying worldview from which the literature and the empirical data are perceived and understood (Col- lin & Køppe, 2003). The research philosophy is of importance to understand beliefs and as- sumptions when developing knowledge as it influences how the research is conducted (Saun- ders, Lewis and Thornhill, 2019). Furthermore, this study takes an interpretive stance as the purpose of an interpretivist research is “to create new, richer understandings and interpretations of social worlds and contexts.” (Saunders et al., 2019, p. 149). In addition, the theoretic perspective of interpretivism seeks to uncover how humans understand and perceive a phenomenon as the paradigm focuses on how humans attempt to make sense of the world around them (ibid). The aim of the study was to comprehend how different interviewees perceive the phenomenon of recommender systems in news media as well as uncover their subjective meanings (about this phenomenon) which aligns with the interpretive belief that a phenomenon should be investigated contextually (Saunder et al., 2019). Furthermore, the ontological as- sumptions in interpretivism acknowledge the existence of multiple interpretations and realities through individuals’ perceptions. The study investigates how the implementation of recommender systems in news media are perceived and understood by individual users taking their different backgrounds into consideration as the ontological assumption acknowledges multiple realities, interpretations, and meanings are socially constructed (ibid). Not only can businesses and management research be complex, but they are often also unique in terms of context which makes the interpretivist standpoint highly appropriate (Saunders et al., 2019).

(21)

3.2 Approach

According to Saunders et al. (2019) researchers can undertake three main approaches to rea- sonings; the inductive-, deductive- and the abductive. This study is in line with interpretivism and therefore an inductive approach is followed. This allows the researcher to collect data to explore a phenomenon and develop findings based on the results of the data analysis (Saunders et al., 2019.) The inductive research approach is closely linked to the chosen research philosophy and contributes to the gaps found in the literature from a user and a company perspective.

By using the inductive approach, the research question is addressed with an explorative and open orientation, but not with what Brinkmann & Tanggaard (2020) refers to as an ‘empty head’. Rich knowledge of the field enables us to ask the best questions (ibid.), thus knowledge of artificial intelligence, news media, recommender systems, and ethics in relation to both have been obtained through the literature review in order to understand the phenomena and detect gaps within the field.

The inductive approach enables the researcher to make sense of empirical collected data through analysis most commonly by formulating theory from the analysis expressed through conceptual frameworks (Saunders et al., 2019). As mentioned above, from the philosophical standpoint this study is interesting in understanding how recommender systems affect the users of news media and their perception of the phenomenon, and the way users interpret their social world. This understanding is exactly developed from an inductive approach. An inductive approach is particularly cornered with the context in which the event of interest takes place (ibid.). Hence a smaller sample of subjects is deemed more appropriate than many subjects often by collecting qualitative data to establish different perspectives on the phenomenon (Saun- ders et al., 2019). In the coming section a description and explanation of how the inductive approach is employed to the research strategy, choice of methods, data collection and data analysis is presented.

3.3 Research Design

The research design is an essential part of the case study as it focuses on answering the research question (Saunders et al., 2019), which for this study is exploring how implementing recommender systems can retain high priority consumers of the Danish news media, Ekstra Bladet.

(22)

From an interpretivist standpoint, an inductive approach to theory was chosen following a methodological choice of a qualitative mono-method study, hence qualitative research and an interpretative philosophy are often connected (Saunders et al., 2019). The following section aims to elaborate on the techniques applied for data collecting and considerations made for sampling. In line with the qualitative research practice and the aim of contributing to existing gaps found in the literature within the field of recommender systems in Danish news media, the empirical basis for this study is an interpretive case study. The case study is based on Yin’s

“Case Study, Research and Applications: Design and Methods” (2018).

3.3.1 Research Strategy

Case study as a research strategy focuses on the “desire to understand complex social phenomenon” (Yin, 2018, p. 4). This empirical method allows the researcher to investigate a phenomenon in a real-world context and to get an in-depth understanding of this phenomenon and its dynamics. This is crucial when conducting a case study as the boundaries between phenomenon and context may not be clearly divided (ibid.).

In this study, the contemporary phenomenon can be characterized in the form of the implementation of recommender systems in news media and the effect on consumers hereof. The case was chosen on behalf of interest within AI and news media which matched the case company, JP/Politikens Hus’ research project Platform Intelligence in News (PIN). The PIN project aims to develop a recommender system for the case company’s online news media Ekstra Bladet, which made a theoretical fit to this field of interest. The tentative research question was defined on behalf of data collected from meetings with the case company and a semi-structured interview with the project manager of the PIN project. A single case study gives the opportunity to analyze the phenomenon of implementing recommender systems in news media and furthermore the effect on consumers which has not been widely investigated in relation to the Danish news media nor the PIN project, in previous work as the case is unique.

3.3.2 Data Collection

In regards to the choice of methods, the primary data source for this study was qualitative semi- structured interviews with stakeholders from the case company and consumers of Ekstra

(23)

Bladet. These interviews allowed the interviewees to elaborate freely on their experience, knowledge, and perceptions (Andersen, 1997). Due to COVID-19, the interviews were conducted by the researcher via Microsoft Teams and recorded video with camera on. This made the interaction similar to a face-to-face interview, besides the disadvantages of not being able to read the non-verbal signs (Kvale & Brinkmann, 2015). All the participants were informed prior to the interview about the purpose of the study and that it would be disclosed due to sensitive information collected from the case company. Thus the thesis is confidential; it enables all the interviewees to speak freely during the interviews (Kvale & Brinkmann, 2018). The interviews lasted an average of 60 minutes, and they were conducted in Danish. To maintain the essence of the interviews the quotations from the findings were carefully translated to Eng- lish afterward. All the interviews were recorded and transcribed with their consent.

3.3.2.1 Meetings and Field Notes

In the early stages of formulating the research question, two meetings were conducted with the Head of the PIN Project, Kasper Lindskow where field notes were collected (Appendix 6). This aligns with the inductive research approach as it allowed the case company to present their knowledge about the project and allowed uncovering of new concepts to be explored instead of confirming existing concepts (Gioia et al., 2013). It was based on these interviews that the exploration of the phenomena was established, the tentative research question formulated, and a baseline for the literature research was found.

3.3.2.2 Semi-structured Interviews

On behalf of the meetings, nine in total semi-structured in-depth interviews were conducted.

Three semi-structured experts' interviews with three stakeholders within the PIN Project and six semi-structured interviews with different users of Ekstra Bladet. Prior to the interviews, the interviewees were sent an email explaining the purpose of the study and the means of confi- dentiality. The format of the semi-structured interviews was created by interview guides (Ap- pendix X-x) to ensure themes were being covered and at the same time making room for unexpected directions while attaining clear structure by following individual interpretations of themes of interest to the researcher (Andersen, 1997).

(24)

As elaborated above, meetings (and an ongoing dialogue about the PIN Project with Kasper Lindskow) were conducted prior to the collection of the three semi-structured experts' interviews. These meetings helped guide the design of the experts' interviews. The format of semi- structured interviews welcomes unexpected directions and at the same time maintains a structure following themes presented in the interview guides. The interview guide was made to ensure that themes were covered (Andersen, 1997) starting with open-ended questions and followed by questions centered around the expert's function within the case company and followed by their knowledge of the PIN Project. The choice of including experts' interviews is based on the premise of exploring in-depth knowledge about the research question and the research object, recommender systems in news media. The interviews provided a thorough understanding of the PIN project including insights from both a business and a technical point of view. The three different experts are presented in Figure 3. While the interviews were limited they were deemed sufficient for the purposes of this study.

3.3.2.3 Motivation Consumer Segment

To ensure that the respondents selected for the thesis suited the consumers of Ekstra Bladet, secondary data from Kasper Lindskow was gathered. In addition information about different consumer segments of Ekstra Bladet was used to assemble the semi-structured interviews (Fig- ure 2). Ekstra Bladet has eight different consumer segments as shown in the Motivation Seg- ment Model hereunder. Hence the implementation of recommender systems are working on their online platform ekstrabladet.dk only the four consumer groups in the middle dark grey area, who are online, were of interest. Though two of these four segments were selected for the thesis as these consumers were of highest priority and interest for Ekstra Bladet. The high priority consumers were The News Junkie and The Larger Consumer. The diversity in gender was equal - three male and three female respondents to explore whether these had any influence.

Educational background and age were also selection criteria (Appendix 2 and Appendix 4).

(25)

Figure 2: The Motivation Segment Model (Appendix 1).

Along with the expert interviews, six semi-structured interviews with users of Ekstra Bladet were conducted in order to uncover the perception of the implementation of a recommender system on Ekstra Bladet’s digital platform from a user perspective. A total of ten consumers were contacted and consented to participate. However, the cases were sampled by the Motiva- tion Segment Model (Appendix 1) within the categories of the two consumer segments; The News Junkie and The Larger Consumer. During the screening, four of the participants did not meet the sample criteria of either consumer segment. Thus a total of six interviews were conducted with consumers reading Ekstra Bladet online falling into the characteristics of either the News Junkie consumer segment or the Larger Consumer segment. The interviews provided an in-depth understanding of the two different consumers segments’ perception of the implementation of recommender systems in news media. Thus the thesis aims to explore how Ekstra Bladet can retain their primary consumers when implementing recommender systems on Eks- tra Bladet’s online platform with the purpose of getting insight into the consumers’ news media habits, their relation to the case company Ekstra Bladet and recommender systems used in news media. The interviews are presented in the following figure.

(26)

Inter- viewee

Titel Interviews themes Length Date

Kasper Lindskow

Head of Pro- ject on PIN and Innova- tion

The PIN Project, serves as an expert to the overall aim of the project and the opportunities and downsides with developing a recommender system for Ekstra Bladet.

Lindskow supports the knowledge about recommender systems from a news media perspective.

63 min. May 11 2021

Ole Sloth Director in Chief

The case company, serves as an expert from a strategic perspective and contributes with in-depth knowledge about the reason- ing of initiating the PIN project from a business perspective.

88 min. May 31 2021

Johannes Kruse

Machine Lear- ning Engineer at Ekstra Bla- det

The development of recommender systems for Ekstra Bladet, challenges, benefits in- depth knowledge about how the recommender systems are being developed and what challenges and benefits the systems has to offer the users.

55 min. June 11 2021

Respondent 1

News Junkie, male, 27, me- dium long education, salaried employee

News media habits, perception of recommender systems in news

77 min. June 13 2021

Respondent 2

Larger Con- sumer, male,

51 min. June 14 2021

(27)

27, salaried employee Respondent

3

Larger Consu- mer, female, 25, graduate

69 min. June 15 2021 Respondent

4

Larger Consu- mer, female, 24, graduate

News media habits, perception of recommender systems in new

55 min. June 17 2021 Respondent

5

News Junkie, male, 27, higher education, salaried employee

39 min. June 17 2021

Respondent 6 (two appendix)

News Junkie, female, 26, student

45 min. June 17 2021

Figure 3: Interviewee summary

As shown in Figure 3 the experts' interviews expanded the knowledge about JP/Politikens Hus and the news media Ekstra Bladet’s relation to the PIN Project and their development of recommender systems. The different consumers from the two consumer segments supported the in-depth knowledge about the consumers’ opinions, concerns and understanding of implementing recommender systems at Ekstra Bladet.

3.3.3 Data Analysis

For this thesis, qualitative data analysis has been applied by the approach of Gioia et al. (2013).

The most common form of analysis today is categorizing interviews to encode qualitative empirical data (Kvale and Brinkmann, 2015). As this study wants to explore the concepts and

(28)

perspectives of recommender systems in news media, the performed coding was detecting words rather than having a linguistic approach to the data collected. By using thematic analysis, the perceptions of the phenomena are explored in a deeper manner ensuring the interviewees’

perceptions and understandings are being captured (Braun & Clarke, 2016). Hence themes and patterns were detected inductively when processing the data to create an understanding of the phenomenon. Coding the qualitative in-depth interviews by a thematic analysis corresponds to the philosophy standpoint of the thesis as it “search for themes, or patterns, that occur across a data set” (Saunder et al., 2019). When analyzing the interviews the thematic analysis has been utilized to make open coding of the data which was done manually by the researcher. Open coding was an approach to identify and describe patterns of semantic content in the data and organize these in the first-order concepts in order to openly investigate what the participants found to be important factors, without being biased by theory. Secondly, these concepts were merged together to make second-order themes. Lastly, these themes were linked together to create three aggregate dimensions concerning Digital News Media, Recommender Systems in News Media and Ethical Considerations. Assessing the second order themes, it is important to notice that they are not necessarily mutually exclusive to the different dimensions, as shown in the Coding Scheme (Appendix 6) but rather have the most influence on the portrayed dimension. These dimensions are developed more in the section of findings.

Figure 4: Data Structure

(29)

So far, an interpretive standpoint has been adopted to research and understand the reality in which a phenomenon exists. An inductive approach to theory development has been selected as the research departs from collected data. A single case study has been conducted to develop in-depth insights into the phenomena and its context. Going forward, the case setting is presented followed by the findings from the data analysis, where relevant theories were applied to interpret the findings, in order to show broader meanings and implications of these (Braun

& Clarke, 2016). Thereafter a discussion of the findings is presented and lastly a conclusion together with further research.

3.3.5 Case Setting

The empirical setting for the case study is the Danish news media company JP/Politikens Hus and their news media Ekstra Bladet. JP/Politikens Hus has together with several Danish uni- versities including Copenhagen Business School (CBS), Copenhagen University (KU), and The Technical University of Denmark (DTU) initiated a research project combining artificial intelligence and news media. The PIN Project stands for Platform Intelligence in News (PIN) and is financialized by the Innovation Fund Denmark and JP/Politikens Hus with nearly 17 million Danish crowns (JP/Politikens Hus, 2020). The project is collecting anonymized data from the Ekstra Bladet’s website to “create a more engaging, more relevant and informing news media by the help of artificial intelligence.” (ibid.). Some of the main stakeholders are Kasper Lind- skow, Head of the PIN project, who is in charge of the project, Anders Søgaard Coordinator for NLP in PIN, Professor of data science at KU who is developing the NLP models in Danish, and Mikkel Flyverbom Coordinator for strategy and ethics in PIN, Professor (MSO) at Department of Management, Society and Communications at CBS who is in line for the ethical goals within the project (Copenhagen Business School, 2020).

The overall vision and purpose for the PIN project is divided into three milestones;

Create a vastly more relevant, engaging, and informing news experience via ethically respon- sible platform intelligence that is geared specifically for news publishing.

Advance State of the Art in Recommender Systems (RS) and Natural Language Processing (NLP).

(30)

Reinvent core aspects of news publishing by enabling publishers to automatically produce different versions of their offerings to specific news users, and to extend the scope of news and service content that can be offered (Sloth, 2021).

The above stated purposes are the long term strategic visions for the PIN Project. For this case study and for Ekstra Bladet, the vision is on a technical level focusing on two main technologies; Natural Language Processing (NLP) and Recommender Systems. NLP is rapidly developing in the field of news media and enables an understanding of language processes (Lindskow, 2021). Recommender systems are based on algorithms making it possible to recommend content such as articles to users online (ibid.). The progress for this vision can be seen in Figure 5 below and involves technical development, live tests, and effect studies on Ekstra Bladet’s digital platform. Both tracks support the identification of strategic and ethical guidelines for the implementation of both NLP and recommended Diffusion Geographic Models (DGM) (Sloth, 2021). These models are capturing geographical distance and consumer preferences.

Figure 5: The Process of the PIN Project

Ole Sloth, CEO at Ekstra Bladet saw potential in the PIN project which from an editor-in-chief perspective is creating a system that creates business value, is ethically justifiable, and aligned with the rules that news media are obligated to follow as part of the press offense (Medieans- varsloven, 2014). Ole has an ongoing collaboration with Kasper Lindskow, Head of PIN and Head of Research- and Innovation Director at Ekstra Bladet, who is a key stakeholder in the PIN project, hence Ekstra Bladet is an important partner and he also has an overview of the bigger picture, timeframe and deliverables. A part of the PIN Project focuses on developing a recommender system for news media that can handle the massive flow of news and

(31)

recommend users better personalized content far better than is done today (Lindskow, 201).

This is what Lindskows aims to fulfill with the PIN project. The recommender systems are being developed as a collaboration between the machine learning engineers at Ekstra Bladet and professors at The Technical University of Denmark. Johannes Kruse is one of the machine learning engineers on the project and has expert knowledge of these systems, their strengths, and weaknesses.

(32)

3.3.5 Delimitation

When evaluating the quality of data three classic evaluation criteria revolve around validation, reliability and generalizability (Kvale & Brinkmann, 2015). Validation refers to the question of the researcher's moral integrity and if the interview investigates what they were meant to (ibid.) In a qualitative study bringing different aspects into the research can create a high valid- ity.

Reliability refers to the credibility of the findings and the consistency thereof. It is connected to the question of other research finding similar biases and information. It can be difficult to replicate the data as it may be subject to change and reflect the reality of the time when it was collected (Saunders et al., 2019). Regarding the different semi-structured interviews collected for this study, the expert interviews could presumably be replicated as they elaborate on their expert knowledge regarding different subjects. However, the interviews with the customer segments would be challenging to replicate as even the same customer segments have very different opinions in relation to the same themes and these might shift over time as the respondents were influenced by their feelings in relation to the topic asked in connecting to what happens at the moment in the society in which they are part of. The interviews for the thesis were conducted to explore a present image of the respondents’ opinions and therefore that is not a concern (Kvale & Brinkmann, 2017).

4. Findings

To answer the research questions, the following chapter presents the empirical findings de- rived from the qualitative data collection. The first section explores the different consumer segments of Ekstra Bladet and their habits in relation to news. The second section investigates the consumers perception of recommender systems in news media and in relation to the case.

Lastly these perceptions include the dimension of ethical considerations in relation to AI and the implementation of recommender systems.

(33)

4.1 Digitalized News Media

All the selected interviewees read their daily news, online. This could either be on a computer, tablet, or phone. Few of them use tabloid news occasionally, but all of them are reading their daily news online. This was a criterion when selecting the respondents as the two segments;

The News Junkie and The Large Consumer are consuming news online (Appendix 1). Further- more, the respondents named several different news media when asked where they read their news. This ranges from the public service media DR and TV2 to Børsen, BT, Politiken, Berling- ske, and Jyllands-Posten. All respondents have in common that they read Ekstra Bladet – some more frequently than others; “I read Ekstra Bladet every day” (Respondent number 1, 2021).

Reading Ekstra Bladet was also a selection criterion when selecting the interviewees for this study.

4.1.1 Consumer Habits

As presented in the Methods section the selection of respondents have been selected by the Motivation Segmentation Matrix (Appendix 1) and focused on two main segments: The News Junkie and The Large Consumer as these are the highest priority segments for the case company and hence the segments they were most interested about.

The Larger Consumer wants to explore the latest news about many different topics from trend- ing tv shows and series to the latest news in society. As this segment wants to be updated all the time and has a high consumption of everything online, their news media habits are not only about breaking news or news concerning society. This corresponds to Respondent 2 who elaborates on his way of consuming news:

“It's a broad spectrum I seek when I read my news. It’s just everything that can fascinate my brain and at the same time everything that can enlighten me about what's going on in the world around me.” (Respondent number 2, 2021).

For the Large Consumer “to be updated” means that they can be a part of the talk when being together with friends, family, or colleagues (Appendix 2). Even though they might not have an interest in the given topic discussed, they still want to read news articles about it to be updated

(34)

and able to participate in the conversation. This is portrayed in Respondent 3’s answer to why and what she reads when reading news articles:

“It may be that someone has played golf. I do not watch golf and I do not care about it. But still, I just read the article. I just like to know what's going on.” (Respondent number 3, 2021)

As the quote illustrates this segment is not limited to news about society or politics, but more or less wants to know ‘everything about everything’. They want to read the news article in depth (Appendix 2), as the headline might show “a slightly misleading picture” and not the whole story of what happened (Respondent number 3, 2021).

The Large Consumer likes both serious and entertaining news content. “It can be a good laugh, and it can also be scientific or enlightening. It can be all of these things.” (Respondent number 2, 2021). As the quote indicates the Larger Consumer does not mind if the news gets funny or

“cheesy” (Respondent number 4, 2021) at times. The news should be easily accessible in their breaks or spare time during the day when they are laying on the sofa or even sitting on the toilet (Respondent number 2,2021). This segment seeks to be entertained and informed (Appendix 2). Also, The Larger Consumer wants more angles on a news story and quick updates while things are happening.

The News Junkies are in contrast to the Larger Consumer mostly interested in news about the society and big current events in Denmark and the rest of the world (Appendix 4). They find their news from several news sources: “Primarily I read my news from news apps, where I have all Danish news apps - in relation to particular politics, society and news” (Respondent number 1, 2021). This segment reads the news to be informed not to be entertained and has an urge to know what is going on in relation to the society they are a part of: “Roughly speaking, I read news almost all the time. Really too often, I think.” (Respondent number 6, 2021). As the state- ment indicates as well as the name of the segment (News Junkie) reading news is done many times throughout the day and also during breaks, hence it is an urge for them to stay updated.

The segment is looking for different angles on a subject which Respondent 6 indicates by saying that she is following people on social media that she agrees with, but also people she disagrees with. The respondent indicates that getting a variety of perspectives on news stories is a posi-

(35)

In addition to the variety of news, News Junkies also tend to have an opinion about what people should read (Appendix 4). Respondent 6 touches upon this tendency when she talks about her parents and their news habit behaviors, where they ‘only’ read Politiken and watch DR and then thinks they are “good to go”. The answer indicates that the segment perceives news media as an important part of being informed and the importance of getting your information from different sources, not just two to be well-informed and oriented from different perspectives. In addition, Respondent 5 contributes to this tendency about having a clear opinion about other people's news reading habits;

“I do not hope that there are any people who get all their information from there [Ekstra Bladet]

because it is often the case that you must be really source critical about what they write.” (Re- spondent number 5, 2021).

The two different segments and respondents have a different approaches to reading news. The characteristics of The News Junkies are that they want to be updated frequently on what happens in society. The Large Consumer wants to be updated on everything the whole time – not limited to politics and news. Both segments have in common that they want to read the news articles in-depth and not just a headliner.

Both segments are aware of Ekstra Bladet’s brand and the fact that they read news on their platform even though they find it unreliable, sensational, or both. Ekstra Bladet is perceived as a “stupid news media” (Respondent 5) by respondent 1 and respondent 5. This perception of news media does not match with the New Junkies self-representation where 25% of the News Junkie segment has a higher education like Respondent 5. This mismatch is also indicated by respondent 1:

“I feel like I would like to keep EB at home, but I also do not want to be him who keeps EB. So that's why I do not have it.” (Respondent number 1, 2021)

Yet Ekstra Bladet is used by this segment as a guilty pleasure and functions as an awry and fun supplement to other news media (appendix). Respondent 5 mentions this by having a subscription on Børsen but uses Ekstra Bladet as an addition to the news he gets from Børsen: “Then

Artificial Intelligence in Digitalized News Media