Language selection

Government of Canada / Gouvernement du Canada

Search

Diversity by design — Diversity of content in the digital age

February 2020

Prof. Dr. Natali Helberger,
Dr. Judith Moeller, Sanne Vrijenhoek

On this page

List of acronyms and abbreviations

ACCC
Australian Competition & Consumer Commission
AI
Artificial intelligence
AIS
Artificial intelligence systems
AVMSD
Audiovisual Media Services Directive
CEFIR
Center for International Relations Studies
EBU
European Broadcasting Union
ECHR
European Convention on Human Rights
EDiMA
European Digital Media Association
ERC
European Research Council
FTC
U.S. Federal Trade Commission
GNI
Global Network Initiative
ICT
Information and Communication Technologies
ISP
Internet service providers
NGO
Non-governmental organisation
OAS
Organisation of American States
OSCE
Organisation for Security and Co-operation
UK
United Kingdom
UNESCO
United Nations Educational, Scientific and Cultural Organization
US
United States of America
WSIS
World Summit on the Information Society

Disclaimer

This document has been prepared for the Department of Canadian Heritage and the Canadian Commission for UNESCO by Dr. Natali Helberger, Dr. Judith Moeller and Sanne Vrijenhoek. The views, opinions and recommendations expressed in this report are those of the authors and do not necessarily reflect the official policy or position of the Government of Canada. Responsibility for any errors, interpretations or omissions lies solely with the author.

Introduction

The digital environment presents citizens and the media with an unprecedented opportunity to access content from a myriad of different voices and outlets, but also with a new challenge of how to make a meaningful selection. Such a selection needs to be responsive to individual information interests, individual constraints in time to read and attention, but also to the broader interests of citizens in being informed and being exposed to a diversity of ideas and opinions that together mark the society in which she is living. Democracies do require that “people in general, and especially differing groups, get to debate their views internally among themselves, receive information relevant to their interests and views, rally support for their group, and finally present their views to the world at large”.Footnote 1 Making such a diverse choice has traditionally been the task of journalists and editors, trained experts in deciding what contents are worth reading and how a sufficiently diverse selection of news should look like. Thanks to technological advances in what Cherubini & Nielsen haven called ‘editorial analytics’, the task of selecting and curating media content is increasingly being passed on to (automated) recommenders. Recommendation algorithms are increasingly being perceived as useful tools to provide the audience with more personally relevant reading suggestions, manage user attention, and unlock the long tail.Footnote 2 Algorithmic recommenders are integrated in social media news feeds but also increasingly added to news platforms as value-added services and critical part of news apps on mobile devices. With the growing trend to algorithmic recommendations, also the question of how to engrain public values, such as diversity, into the algorithmic design, is gaining in vital importance for digital democracies.

The task of data-driven recommendations is to filter the growing abundance of information online. Based on data of what people like to read, what their friends like to read, what contents sell best, what contents sell less, recommenders use machine learning and AI to make ever smarter suggestions for us.Footnote 3 Generally, four types of news recommender algorithms can be distinguished. Recommenders can be distinguished according to the input data they rely on. Some recommender algorithms make personalized recommendations on the basis of metadata (content-based), insights into what other users like to read (collaborative filtering) data on their users (knowledge-based), or a combination thereof.Footnote 4 Another important distinction is that between self-selected recommendations (here the users determine the selection criteria and feed the system with their preferences) and preselected recommendations (here the media determine the selection, based on volunteered or inferred data).Footnote 5

Recommenders can be optimized for different goals. Probably most widely-spread are recommendation algorithms that are (in all likelihood) optimized for short term metrics such as user engagement, time-spent, advertising revenue, clicks, and more recently: more contents from friends and family and ‘user happiness’. These can be found e.g. on social media sites like YouTube or Facebook. The focus on short-term metrics resulted in the well-known concerns about the medium-term effects of such choices, which are feared to favor polarization, filter bubbles and echo chambersFootnote 6 and an overall decrease in the diversity of ideas and voices that the audience is being exposed to.Footnote 7 But more recently, a trend could be observed towards more value-sensitive recommender design, particularly in the quality media that are also increasingly embracing and experimenting with (personalized) recommendations. As the Frankfurter Allgemeine Zeitung informs its users: „Nicht vom Gleichen immer mehr, sondern intelligent ergänzt mit überraschenden Inhalten, die Ihr Leseerlebnis bereichern sollen. Das Angebot soll so vielfältig sein wie Ihre Leseinteressen… Damit Sie Ihre Zeit dem Lesen und nicht dem Suchen widmen können,“[“Not more of the same but instead intelligent recommendations that surprise and enrich your reading experience. The news offer should be as diverse as your reading interests.....so that instead of searching for news, you can invest your time in reading“]. And the Neue Züricher Zeitung wants to find “clever new ways to integrate personalized experiences to the core of our newsproduct – with editorial integrity in mind”.

Unwanted filter bubbles are bad recommender design. A growing body of, on the one hand, practice examples in the media and, on the other hand, academic literature explore the potential for ‘diversity by design’.Footnote 8 And yet, translating a concept that is as ill-defined and malleable to normative interpretation as media diversityFootnote 9 into concrete metrics and measures that can inform algorithmic design is not a trivial task. Too often, media diversity is equated with simplified statistical measures of similarity or some measure of serendipity.

As part of their engagement strategy on Diversity of Content online, the Canadian government has asked us – based on our expertise and long-time involvement in research on diversity by design to provide a thought leadership paper on diversity by design. More concretely, we were asked to: “provide a more thorough analysis of that selected action [the optimization of algorithms to ensure access and exposure to diverse content], including a feasibility assessment and important considerations, i.e. what elements or concepts would be important to take into account by stakeholders, from a global, multi-stakeholder perspective.”

This paper is divided into four sections. After the introduction (section 1), we will briefly introduce the concept of (exposure) diversity, some reflections on the importance of exposure diversity on democratic resilience of citizens as well as a brief overview of current insights into the effects of news recommender algorithms on exposure diversity (section 2). In section 3, we will then briefly outline the concept of diversity by design, and give a brief overview of the literature and attempts to translate diversity, as a normative notion, into metrics that can inform recommender design. The next section (section 4) will then offer some insights on the steps that need to be taken to make diversity by design feasible. This section is informed by the current state of academic literature and most importantly by our research into the topicFootnote 10 as well as the insights from a Dagstuhl seminar that we organized, together with an interdisciplinary team of leading thinkers in this area in last November 2019 at Schloss DagstuhlFootnote 11. The objective of Dagstuhl Perspectives seminar was to develop a broader vision of, and recommendations for the future of (diverse) news personalization recommenders as the result of a collaborative effort by different areas of academic research (media studies, computer science, law and legal philosophy, communication science, political philosophy, and democratic theory). In so doing, the insights from this seminar are of direct relevance for this paper.

Exposure diversity as a concept

Brief introduction into exposure diversity as a normative, inherently democratic concept

“There can be no democracy without pluralism”.Footnote 12 Diversity is a normative value and has a long tradition in free speech and democratic theory.Footnote 13 Democracies do require that “people in general, and especially differing groups, get to debate their views internally among themselves, receive information relevant to their interests and views, rally support for their group, and finally present their views to the world at large”.Footnote 14 Media diversity is instrumental in achieving that goal.

Media diversity is a central value in media law and policy, and yet it is surprisingly ill-defined.Footnote 15 The Council of Europe, an international institution that has had a significant impact in interpreting the European Convention of Human Rights, and influencing European media law and media law concepts defined diversity originally as a “variety of media content reflecting different political and cultural views […] made available to the public”.Footnote 16 From the point of diversity by design this is still a very generic, unspecified definition that still leaves many questions open, and also later versions of the definition were not able to make the notion more concrete. Having said so, from the Council of Europe perspective it does at least become clear that diversity is not a goal in itself, but instrumental to “promoting a critical debate and wider democratic participation of persons belonging to all communities and generations”.Footnote 17 What also becomes clear from the earlier Council of Europe definitions of diversity is that, for a long time, the primary focus of conceptions of media diversity were on diversity of supply, that is: the diversity of information that is available in the news media. In the digital environment, however, where the available information offer by far exceeds the number of information citizens can consume, the question of exposure diversity is gaining in importance and becomes increasingly pivotal for the realization of the public values that media diversity serves in the first place.

Napoli defined exposure diversity as “the diversity of content or sources consumed by audience members, which, of course, maybe very different from the diversity of content or sources available.”Footnote 18 Exposure diversity is the idea that for diverse media content to have an effect on the way people engage with information, each other and the public debate, it is not enough to concentrate on the diversity that is supplied, but also the amount of, and the conditions under which users (can) consume diverse content. To the extent that the amount of information supplied exceeds the amount of information users can reasonably access and read, the aspect of exposure diversity is gaining in importance for the realization of the public policy goals that the concept serves. The purpose of algorithmic recommenders is to filter and sort the amount of information available and make (personalized) selection. As such, recommenders have a critical task in impacting exposure diversity, and the public policy objectives behind diversity, as also acknowledge in more recent media policy initiatives, including of the Council of Europe.Footnote 19

Factors and likely consequences with respect to the democratic resilience of citizens

Concerns about limited exposure to diverse information online are frequently associated with two expected consequences for democratic societies. First, filter bubbles are seen as one of the key drivers of polarizationFootnote 20 Footnote 21, = The underlying reasoning of this argument is that exposure to counter-attitudinal views in the media should lead to increased understanding for the opposing side and hence fosters tolerance. As algorithms detect preferences and ideological leaning, they are expected to reduce the opportunity of citizens to engage with such content in a way that is unobtrusive to the user of the platform Footnote 22. The other potential consequence of reduced exposure diversity is related to the democratic function of the mass media to set a shared public agenda. Algorithmic filter systems can potentially amplify citizens’ existent interests in specific topics, by filtering out information about other topics, including those at the top of the public agenda.Footnote 23

Impact of algorithms on exposure diversity

Empirical research into the likely effects of algorithms is limited by the availability of data. Large scale studies about the impact of algorithmic filtering on the exposure to (non-)diverse media content, and its consequences on civic attitudes, knowledge and behavior would require access to individual-level data provided by platforms, which is currently not available Footnote 24. A notable exception is one study conducted by researchers at Facebook that found that the use of the platform leads to more diverse exposure.Footnote 25 There are, however, several empirical studies that provide insights into the question to what extent citizens encounter more or less diverse news in algorithmic systems and a few experimental studies on the impact of potentially limited exposure diversity.Footnote 26 Fletcher and Nielsen demonstrated in a comparative study based on representative survey data from four countries, that incidentally receiving news through platform provided algorithmic filter systems (Facebook, YouTube, Twitter) by and large leads to a larger diversity in sources, especially among younger people and those with lower interest in the news.Footnote 27 This is in line with our research which demonstrated that different types of algorithmic filtering can be even more diverse than human editors in a data-scientific experiment based on real-world data of one of the largest Dutch newspapers.Footnote 28 These findings are echoed in the work of other scholars who studied exposure diversity in search results. Footnote 29

Even though we might be seeing more diversity when we use social media and search, it's possible that this diversity may consist of more partisan or polarizing news sources. A study published by a team of researchers in the US looked at people's exposure to content shared by users aligned with a different political ideology on Twitter.Footnote 30 If they were Republicans, they received more messages from Democrats, and vice versa. The researchers measured attitudes before and after this process and what they found is that as people paid attention to messages from the opposing side, their attitudes began to polarize, and they became more entrenched in their original beliefs.

Diversity by design

Brief introduction into the concept

Diversity by design is a form of value by design. Value by design has been defined as “a theoretically grounded approach to the design of technology that accounts for human values in a principled and comprehensive manner throughout the design process.“Footnote 31 The value at stake here is exposure diversity, as defined in section 2, and what characterizes attempts to design for (exposure) diversity is what Friedman et.al. called a “pro-active design component”. Footnote 32 A pro-active design component means finding ways to conceptualize and implement exposure diversity in a way that can inform the design of recommendation algorithms. This is distinct from research that seeks to measure (empirically) the effects of news recommenders on the diversity of contents that users are exposed to.

Translating a value into concrete design requirements is not a straight forward process. As Friedman et.al explain, it can involve going through several iterations of testing and adjusting the parameters or models that inform an algorithm.Footnote 33 Van der Poell distinguishes at least three steps in such a process:Footnote 34

First: conceptualizing the value (including an exploration of the reasons why a particular value is considered valuable)

Second: translation of a general value into one or more general norms [or metrics];

Third: The translation of these general norms into more specific design requirements.Footnote 35

While much of the existing work to translate diversity into design is focused on the recommendation algorithm alone, a review of the theoretical diversity by design literature reminds us that we cannot see the algorithm as the design object in isolation, but also need to look at the broader societal context in which the algorithmic is being implemented, including the broader cultural context,Footnote 36 but also an investigation of the affected stakeholders, and how they negotiate values, as well as the direct “human context in which the technical artifact is situated”.Footnote 37 News recommenders are an excellent example to explain why diversity by design can hardly be achieved without taking into account the broader human context, here: the news business, in which the algorithm operates. There are several reasons for this. First, the possibility of exposure diversity depends on the availability of content in the pool. If the quality and diversity of the pool is low, recommenders have insufficient options to provide good recommendations. That means exposure diversity ultimately is dependent on external diversity. Secondly, research into algorithmic bias showsFootnote 38 that often human bias is amplified in algorithmic systems. This pertains not only to the training data used run AI systems but also to the ability of those designing the system to observe and interpret bias. Third, while designing the algorithm is often a task of the technical department, it is the editorial team that determines which selection of contents conforms with the editorial policy, commitment to diversity, the mission of a particular news outlet, etc. And yet, in reality, there can be a disconnect between these two departments. In the worst case, there will not be any communication at all.Footnote 39 Fourth, as the normative expectation of what “good diversity” is a moving target, it is important to consider the process that leads to the definition of the level and kind of diversity that should be achieved. This way systems that are diverse by design remain transparently adaptive to societal and normative change.

Finally, it is important to be mindful of another lesson from the general diversity by design literature, namely that there are also certain limits to value sensitive design, and for our case: the extent to which diversity as a normative concept can be operationalized in concrete recommender design.Footnote 40 This can have to do with the sheer difficulty of translating certain aspects of diversity, or the trade-offs between values that optimizing for exposure diversity can involve, obstacles in the broader socio-economic context in which recommenders are developed (e.g. commercial constraints and need to optimize for profit rather than for diversity metrics) but also the limited effectiveness of recommenders in actually steering user choices. As Friedman et.al. explain: “a given technology is more suitable for certain activities and more readily supports certain values while rendering other activities and values more difficult to realize.’Footnote 41

Brief introduction into the current state of literature on diversity by design

The main focus of the diversity by design literature can be found in the field of computer science. In the computer science literature generating a recommendation is seen as a reranking problem. Given a set of items, the goal is to present these items in such a way that the user finds the item he or she is most interested in at the top, followed by the second-most interesting one, etcetera. However, the problem lies in how to approximate this user interest. Content-based approaches look at the type of items that the user has interacted with before and recommend similar ones. Here one could think of finding topics or overall texts that are similar to what is in the user’s reading history. On the other hand, in collaborative filtering approaches, the algorithm considers what other users similar to the user in question have liked, and recommend those. Most state-of-the-art systems are hybrids of these approaches, but also include elements of serendipity, which facilitates exploration. Typically, this means that recommenders include random content to generate new information about the user. If the user engages with the additional content, the system becomes aware that the current personalization is erroneous and learns what kind of content the user would like to have received.

However, since most optimization metrics are based on a form of similarity, it becomes likely that at some point only the same types of content are recommended. To tackle this, the concepts of novelty and diversity are introduced, which are strongly linked.

In the most commonly cited work on this topic, Vargas et al define novelty as “how different it is with respect to what has been previously seen”, whereas diversity refers to a set of items, and “how different the items are with respect to each other”.Footnote 42 In other words, novelty compares the current set of recommended items to what the user has seen before, whereas diversity compares the differences within that set. The challenge here then lies in how to define this difference or distance. This is often done by considering the respective ‘topics’ of items,Footnote 43 which is used as a common reference, calculates the distance between topics assuming a taxonomy of topics already exists. This does not consider the nature of news items, where stories may span multiple days and new topics may arise every day. It also leaves no room for differing opinions and viewpoints about each topic, which may be relevant for the overall diversity. Another concept related to diversity is that of the ‘long tail’ - do we only recommend the most popular items that we know many people are interested in, or do we also recommend items that perhaps not that many people have had a chance to interact with yet?

In a recent overview of the 'current state' of diversity,Footnote 44 the authors constructed a taxonomy of definitions of diversity. The first decision that needs to be made is whether to calculate individual diversity (on user basis) or on aggregate diversity (on set basis). When focusing on individual diversity a choice needs to be made about the method of feature representation and within which space these features are considered. The representation can be both explicit (e.g. through metadata) or implicit (learned in some way). Current work has mostly focused on explicit features, such as popularity, publication time, source, or sometimes category. In terms of feature space, there can be both pairwise (comparing one recommendation with another) and set-level measures (saying something about the set of all recommendations).

As already becomes apparent from the overview above is that although there are various attempts to conceptualize (exposure) diversity in the computer science literature, these definitions depart from the idea of diversity as some form of variance or dissimilarity. They are not grounded in, and do not refer back to the more normative understanding of diversity (or pluralism, a notion that is often used interchangeably) in the media law, fundamental rights law democratic theory and media studies and communication science literature.

There have been some attempts to measure (exposure) diversity, without, however, attempting to translate these insights into design criteria for algorithmic news recommenders. Defining and measuring media diversity, for example, is an important task of media regulators and supervisory authorities. Their task is to guard over the diversity of media markets. Various measurements have been suggested and tried, including the FTC’s controversial Diversity Index, the UK’s public interest test or Germany’s audience share model Footnote 45 Most of these measures concentrate in one way or other on observing the diversity of sources in a media market while abstaining from statements of when a diversity of sources is diverse enough, or how much (exposure) diversity is necessary so that a heterogenous audience is indeed able to engage in critical debate and wider democratic participation. Until now, probably the most extensive effort to define and measure media pluralism was the European Media Pluralism Monitor. The European Media Pluralism monitor lists more than 300 risk indicators of media pluralism and diversity.Footnote 46 This risk-based approach towards defining diversity was neither directly geared towards assessing algorithmic recommendations, not is it particularly useful in conceptualizing what diversity in recommendations is because of its focus on European media markets in their entirety. Assessing the diversity of the results of personalized recommendations requires an assessment at the (internal) level of the recommendation, or respectively the output of a personalized recommendation. One important insight from the Media Pluralism Monitor Project for this paper, however, it is that diversity cannot be pinned down to one or a few criteria.

Another stream of research into the nature of diversity tries to understand more from an empirical perspective, as the impact that news recommenders can have on diversity (including the question of whether or not filter bubbles exist).Footnote 47 Most of these studies so far benchmark and measure exposure diversity in terms of exposure to counter-attitudinal or counter-ideological content.Footnote 48 This is a relatively straightforward measure of exposure diversity, but also a rather limited one. It may make more sense in the US context that is dominated by a two-party system (liberal vs conservative),Footnote 49 but even in that context, diversity is reduced to whether one encounters content that one agrees or disagrees with.

So far, there are a number of projects emerging that seek to bridge the gap between the empirical, the computer science and the normative perspective on exposure diversity in general, and the diversity by design in particular. There is for example the EBU Peach Project,Footnote 50 the NewsDNA project in Ghent,Footnote 51 Project Diamond in Leuven,Footnote 52 Project Ensure in Delft,Footnote 53 and work done at the University of ZürichFootnote 54 that focus on various aspects of diversity in the (new) media. Maybe the earliest and most systematic project so far (at least to the knowledge of the authors) that seeks to bring together the normative perspective, the computer science/AI perspective and the communication science perspective is the Unlocking News Recommenders-Project at the University of Amsterdam. The project has been funded by the SIDN funds, received support from the PersoNews ERC project and is a collaboration between academics and practitioners (Blendle, RTL). Footnote 55 Unlike the other projects the Unlocking News Recommenders Project approaches the challenge of developing diversity metrics by returning to the origins and normative underpinnings of why we valuate media diversity in a democracy. In the following section, we will briefly outline the main insights from this project.

Bridging the different disciplinary approaches: news recommender diversity from the perspective of democratic theory

One key insight from the definitions of the concept of diversity by the Council of Europe is that diversity is not a goal in itself, but a concept with a mission. We value diversity not because it is ‘sells to be diverse’, but because diversity is considered pivotal in promoting the values that define us as a democratic society. The media diet that people are exposed to should reflect, in one way or another, the diversity of voices and ideas in a society. Diversity, in turn, is considered instrumental in realizing a whole range of goals that we value in a democratic society: from stimulating informed citizens and open-mindedness, to tolerance, cultural inclusion and equal opportunities. The media have traditionally played a pivotal role in the realization of these values, as a platform and communicator of diverse media content. The exact meaning of diversity in recommendations depends at the democratic theory one follows, but also the values that are central in the editorial mission of a particular news outlet.Footnote 56 In the following, we will present four different approaches to conceptualizing diversity in news recommendations, each following from a different democratic theory and placing different public value central. This not to say that one theory or approach is superior to the other. A key message of this section is that in order to understand and conceptualize diversity in news recommendations, a decision is needed about the goals, mission and public values that a particular news outlets pursues. Some approaches or conceptions of diversity better fit the quality media or public service media, others are more compatible with online media or social networks. Each conceptualization of diversity in news recommendations is the result of a balancing exercise and prioritizes certain values over others.

A liberal approach to diversity in recommendations

In liberal democratic theory, individual freedom, including fundamental rights such as the right to privacy and freedom of expression, dispersion of power but also personal development and autonomy of citizens stands central. Unlike under more deliberative or even critical theories of democracy, the liberal model is in principle, sympathetic to the idea that algorithmic recommendations are tools to enable citizens to further their autonomy and find relevant content. The underlying premise is that citizens know for themselves best what they need in terms of self-fulfillment and exercising their fundamental rights to freedom of expression and freedom to hold opinions, and even if they do not, this is only to a limited extent a problem for democracy. This is because the normative expectations of what it means to be a good citizen are comparatively low and there is a strict division of tasks, in which ‘political elites […] act, whereas citizens react’.Footnote 57 As Ferree et. Al. Footnote 58 have put it so placatively: “Citizens need policy makers who are ultimately accountable to them but they do not need to participate in public discourse on policy issues. Not only do they not need to, but public life is actually better off if they don’t.”

Under such a more liberal perspective on democracy, diversity in recommendations would entail a more organic, user-driven approach to diversity – trying to reflect the different interests and preferences of users (few users will only want to watch cat videos and celebrity views). Accordingly, characterizing features of a liberal recommender could be: more prominence for more prominent topics, displaying the main political issues, particularly during election time but else little distance from personal preferences. Instead, offering users the choice between different formats, topics, genres, sources or even between different recommendation logics. There should be possibilities for active user curation as they are autonomy enhancing, but also a high level of respect for personal privacy preferences and personal autonomy (in other words: minimal nudging).

A participatory approach to diversity in recommendations

An important difference between the liberal and the participatory model of democracy is what it means to be a good citizen. Under participatory conceptions, the role of (personal) freedom and autonomy as an essential ingredient for furthering the common good, rather than personal self-development: “[F]freedom is marked by the ability to participate in the public sphere, by the subordination of egoistic concerns to the public good, and by the subsequent opportunity this creates for the expansion of welfare, individual and collective”Footnote 59 Citizens cannot afford to be uninterested in politics because they have an active role in helping the community to thrive: “The stronger civil society is, and the more social capital a society has, the more democracy thrives”.Footnote 60 Accordingly, the media, and by extension news recommenders must do more than to give citizens ‘what they want’, and instead provide citizens with the information they need to play their role as active citizens.Footnote 61

Participatory recommenders have an important role in furthering participatory values, such as inclusiveness, equality, participation, tolerance, active citizenship, etc. which involves a far more principled understanding of ‘participatory or representative diversity’. Instead of simply giving people what they want (at this particular moment), a participatory recommender must pro-actively address the fear of missing out on important information and depth, and the concerns about being left out. The particular challenge will be to make a selection that gives a fair representation of different ideas and opinions in society, while also helping a user to gain a deeper understanding, and feeling engaged, rather than confused. Ferree et al. 2002, speak of ‘empowerment’: “The argument that public participation transforms individuals into engaged citizens implies that media content should first and foremost encourage empowerment” (stress by original authors).Footnote 62 To be truly empowering, media content needs to be presented in different forms and styles.Footnote 63 By extension, this means that diversity is not only a matter of the diversity of content, but also of communicative styles.

What would then characterize diversity in a participatory recommender are, on the one hand, active editorial curation in the form of drawing attention to items that citizens ‘should know’, taking into account inclusive and proportional representation of main political/ideological viewpoints in society; a focus on political content/news, but also: non-news content that speaks to broader public and, on the other hand, a heterogeneity of styles and tones, possibly also emotional, empathic, galvanizing, reconciliatory.

A deliberative approach to diversity in recommendations

The participative and the deliberative models of democracy have much in common.Footnote 64 Also, in the deliberative or discursive conceptions of democracy, community and active participation of virtuous citizens stands central. One of the major differences is that the deliberative model involves a process of actively comparing and engaging with others, also contrary and opposing ideas.Footnote 65 Political and public will formation is not simply the result of who has the most votes or ‘buyers’, but it is the result of a process of public scrutiny and intensive reflection.Footnote 66

This epistemological shift from information to deliberation has obviously also important implications for the way the role of news recommenders can be conceptualized. Diversity in the deliberative conception has the important task of confronting the audience with different and challenging viewpoints that they did not consider before, or not in this way: because the aim of the deliberative process is to broaden the participants' information and enable them to discover their own preferences, that process requires a multiplicity of points of view and/ or arguments. As the individual listens to arguments formulated by others, he broadens his own point of view and becomes aware of things he had not perceived at the outset. Deliberation requires not only multiple but conflicting points of view because conflict of some sort is the essence of politics.”Footnote 67

Concretely, this means a deliberative recommender (or recommendation) to be diverse should include a higher share of articles presenting various perspectives, diversity of emotions, range of different sources; it should strive for equal representation, including content dedicated to different ethnic, linguistic, national groups, as well as on recommending items of balanced content, commentary, discussion formats, background information; potentially some prominence for public service media content (as the mission of many public service media includes the creation of a deliberative public sphere), as well as a preference for rational tone, consensus seeking, inviting commentary and reflection. Social media elements but also actively nudging to consume news content this citizen should know or are unlikely to encounter for themselves could also be elements of a diverse recommendation.

A critical approach to diversity in recommendations

As popular as the deliberative model of democracy is, especially among academics, as many critics did it attract. A main thrust of criticism is that deliberative theory is too much focused on rational choice, on drawing an artificial line between public and private, on overvaluing agreement and disregarding the importance of conflict and disagreement as a form of democratic exercise.Footnote 68 To the contrary, conflict, disagreement and the presence of marginalized voices do not conform to the high standards of democratic deliberation, either because their voice is not public enough, or because it is too shrill, irrational or provocative. Accordingly, critics of the deliberative or participatory model of democracy argue that the focus on reason and tolerance muffles away the stark, sometimes shrill contrasts and hidden inequalities that are present in society, or even discourage them from developing their identity in the first place.

Arguably, under a critical model, news recommenders will not seek to reflect the full diversity of ideas and opinions in society, but instead focus on these marginalized or minority voices and actively nudge users to experience otherness. The critical recommender is aimed at promoting citizenship maybe not even so much but informing citizens about important political topics and events, but on matters of everyday life, where marginalization and exclusion of minorities is particular pressing. This could also imply that constructionist recommenders would also offer room for alternative forms of presentations: narratives that appeal to the ‘normal’ citizen because they tell an everyday life story, emotional and provocative content, even figurative and shrill tones - all with the objective to escape the standard of civility and the language of the stereotypical middle-aged, educated, blank white man.Footnote 69

It is exactly this process of stereotyping that is a key danger of recommenders from the perspective of more critical perceptions of democracy: by using standardized profiles as the basis for recommendations, recommenders can have potentially the effect of re-enforcing biases and muffling away other, more controversial and hence less easily shared aspects of our identity. Transparency about the profile and the ability to adjust would seem important pre-conditions for the democratic role of the constructionist recommender. And even more so than under the deliberative perspective it is important to explain why we are recommended certain items, but maybe even more importantly: who and what has been left out.

Based on this analysis,Footnote 70 we developed a first indicative list of criteria or metrics that could characterize different recommendation algorithms, depending on the values and democratic theory that the recommender is being optimized for. Which values those are will ultimately depend on where the recommender is used. While the liberal recommendation logic can be encountered for example on social media websites and many commercial media websites, the participatory or the deliberative model is more likely to be found on e.g. the websites or news apps of quality media, whereas the critical model is unlikely to be found in a market economy but could inform e.g. the recommendation algorithms of the public service media.Footnote 71

Table 1: four types of democratic recommenders
Recommender Liberal Participatory Deliberative Critical
Values to optimize for Autonomy, self-development, dispersion of power Inclusiveness, participation, active citizenship Deliberation, tolerance, open-mindedness, public sphere Including marginalized voices, defy prejudices
Characteristics

More prominence for more prominent topics

Main political issues, particularly during election time

For the rest: little distance from personal preferences

Choice between different formats, topics, genres, sources

Choice of recommendation logics

Expert recommenders

Tool of active editorial curation, drawing attention to items that citizens ‘should know’

Activating

Inclusive and proportional representation of main political/ideological viewpoints in society

Focus on political content/news, but also: non-news content that speaks to broader public

Background information

Political advertising

Higher share of articles presenting various perspectives, diversity of emotions, range of different sources

Equal representation, including content dedicated to different ethnic, linguistic, national groups

Balanced content, commentary, discussion formats, background information

Prominence public service media content

Personalised nudging to consume news content this citizen should know

Prominence for less popular content, minority and marginalized voices

Actively nudging

Critical tone

Content that is purposefully provocative, opposes, challenges
Form & presentation Possibilities for active user curation, autonomy enhancing, no nudging, privacy friendly design, data portability Accessible, multi-platform, heterogeneity of styles and tones, can be emotional, empathic, galvanizing, reconciliatory, transparency & accountability Preference for rational tone, consensus seeking, inviting commentary & reflection, social media elements Heterogenous, narratives, preference for contents that are affective, emotional, provocative, figurative

In a follow-up step, and with the support of the SIDN funds, the authors of this paper, at the University of Amsterdam, and in collaboration with Dutch media company RTL Nieuws, have started to translate the metrics developed above into a tool that aims to measure diversity in news recommendations. With the tool, news organizations can load their recommendations and evaluate them on a number of different aspects of diversity.Footnote 72 By comparing these results to those of a number of simple baseline recommendation approaches, the effect of the news organization’s own system can be evaluated. Work on the project is still ongoing, with the development of better and more sophisticated metrics, and the implementation of the tool with a range of industry partners.

The road ahead: What needs to happen to make diversity by design work

Diversity is a key public value in democratic societies and recommendation algorithms that sort and order to abundant digital information offer are here to stay. Consequently, it is of paramount importance to promote initiatives that develop smarter recommendation algorithms. Smarter recommendation algorithms should not simply optimize for clicks and short term metrics, but be able to accommodate and promote public values, such as diversity Making recommendations smarter and more diverse needs to become a focus area of media innovation but also public media policy.

Realizing and operationalizing diversity by design is a truly interdisciplinary effort. For some time now, the questions of what diversity is, and how it can be defined followed separate paths in the social sciences, humanities, and in computer science. At the same time, there is a growing awareness that to arrive at meaningful diversity metrics, media studies, communication science, law, and fundamental rights expertise, user-computer interaction studies and computer science need to intensify their cooperation. The Dagstuhl seminar “Diversity, Fairness, and Data-Driven Personalization in (News) Recommender Systems”Footnote 73 gathered experts from a range of disciplines (computer science, law, communication science, political science and sociology) to engage in joint brainstorming of the steps that need to be accomplished before realizing diversity by design. The seminar was an example of the growing acknowledgment of the interdisciplinarity of this challenge, but also: an example of how one of the main obstacles in this respect – the different traditions, methods, and languages in the different disciplines can, and already is being overcome. Many more truly interdisciplinary and structural exchanges such as the Dagstuhl seminar need to happen to allow us to re-conceptualize diversity as a normative and democratic notion in a way that can lend itself to formalization in algorithmic design. This is also why one of the recommendations of the Dagstuhl seminar was the creation of an international, interdisciplinary research initiative with a joint lab to spearhead the needed interdisciplinary research needed, boost practical innovation, develop reference solutions, and transfer insights into practice. This initiative and its lab must combine the best (inter)national expertise from fields like computer science, social and behavioral sciences, and law, as well as industry and regulators, to ensure diverse, transparent and explainable news recommendations.

The question of how to realize diversity by design is not only an academic question, it is a societal and practical question. Insofar, academics, as well as media professionals and society in more general, share an interest in finding ways to more innovative and value-sensitive forms of deploying recommendation algorithms. To make diversity by design work on the ground, and for society, academics need to work with the media industry on various levels: on the conceptualization of diversity, the formalization of diversity, reiterative testing and measuring and understanding the implications of different recommendation logics on exposure diversity. Regarding the conceptualization of diversity, it is important to acknowledge that while diversity is a key public value, it is not an absolute value. Ultimately, the optimal level of diversity is the result of a balancing exercise between the societal benefit of being exposed to diverse content in the media, and other, potentially competing values, such as privacy,Footnote 74 autonomy and the right to informational self-determination, but also economic rights, such as the right to economic sustainability and to operate a business, but also: the interests of users in simply reading relevant content, or being entertained.

Too much diversity can backfire and have a negative, rather than positive effect on users.Footnote 75 First experiences teach us that defining and formalizing diversity is moreover a very resource-intensive effort, so resource-intensive that it might go beyond the (often) scarce capacities, and daily priorities of news media. Here, the media industry can clearly benefit from collaborations with academia. Ultimately, to do so in a meaningful way is not a one-time off exercise (see also section 3.a), but a process that involves real-life testing and re-adjusting, and doing so in reiterations. This is where academia, in turn, benefits from and depends on cooperation with the media industry. Finally, the growing awareness of the importance of diversity by design is promising, but should not distract from the fact that in the digital society, recommendation algorithms can be powerful instruments with potentially far-reaching implications for the state of the democratic debate, the meaning of informed citizenship, the distribution of political and economic power, and democratic oversight. This is why next to finding ways to diversify recommendations, independent oversight and monitoring is needed of the way recommendation algorithms affect overall exposure diversity and the quality of the public sphere. Also, this is an aspect of diversity by design – to develop metrics that allow measuring the impact of recommendations on exposure diversity. To do so, academia and regulators not only depend on collaboration with the industry in designing and testing such metrics but also on access to data on the impact of recommendation algorithms, for example on social media platforms. Sometimes this information is being shared voluntarily by media industry actors because own intrinsic motivations to understand the impact of their recommendation systems, but often access to data is still subject to trade secrecy and proprietary contractual conditions. This is why another recommendation of the Dagstuhl seminar concerned the need to create rights and procedures for data sharing – a claim that is also increasingly understood and echoed in policy debates in Europe.

As explained earlier (section 3.a), diversity by design goes beyond the actual recommendation algorithm. Diversity by design also needs to take into account the broader socio-technological context in which algorithms operate, the different stakeholders involved in designing, deploying and shaping algorithms (from editors to marketing directors, journalists, platforms, designers, and users). A recommendation is the result of an algorithmic feedback loop that is shaped by the input from all these actors, and the choices that they make. As the Council of Europe in its declaration on the manipulative capabilities of algorithms declared: “When operating at scale, such optimization processes inevitably priorities certain values over others, thereby shaping the contexts and environments in which individuals, users and non-users alike, process information and make their decisions. This reconfiguration of environments may be beneficial for some individuals and groups while detrimental to others, which raises serious questions about the resulting distributional outcomes.”Footnote 76

Within newsrooms, it is important to realize that the decision of what conceptualization of diversity and choice between competing values to optimize for is essentially an editorial decision. This is a decision that must be informed by the editorial mission of a particular outlet, and can also vary between different outlets. The kind of diversity that a public service media operate may decide to optimize for can be different from the kind of diversity that the online version of a commercial medium, the online version of a tabloid press or a social network may optimize for. On the contrary, the decision of how to implement these priorities in algorithmic design is a decision that will be taken by technical departments. In reality, the necessary dialogue and interaction between the editorial and technological departments are not always taking place yet. This is why we have argued elsewhere that journalistic ethics require a process of re-thinking also of the internal routines, the need for creating procedures to enable what needs to be a collaborative process between editors, marketing and designers, but also: some formalization of the mission statements, goals and objectives that inform this process to enhance public accountability.Footnote 77

Another perspective that yet deserves more attention is the perspective of the user. Ultimately, the goal of diversity by design is to improve the chances of users to encounter diverse content, also in an algorithmically curated environment. Having said so, whether (diverse) recommendations will ultimately do so or not may not only be a matter of the algorithm itself, but also the way the recommendations are offered to the user, the trust that users have in a media outlet, and by extension: its recommendations, but also the purpose for which the user is searching for information, maybe the timing as well as the degree of personal agency that the recommender allows users to have, and that users are willing to exercise. As said earlier, recommendation algorithms do not operate in isolation, but recommendations are also the result of the interactions with users. As a result of personalization and the proliferation of data-driven recommendations, the relationship between the media using these algorithms, and users is changing, and is becoming more interactive, immediate, personal and potentially pervasive. Insofar, diversity by design not only involves the formalization of a normative value into algorithmic design but also: re-thinking the optimal role and responsibilities of the media vis-à-vis the audience.

Media organizations themselves do not operate in isolation, but in a highly competitive environment where the legacy media is competing with social media, search engines, news aggregators and other alternative media platforms for the (scarce) attention of users. Any attempt to implement and operationalize diversity by design by e.g. the legacy media must also be seen in the light of this competition, as it will influence the way public values (such as exposure diversity) will be weighed against other values, such as the need to generate income. Much will depend on whether offering diverse and high-quality recommendations will be perceived by users as a value-added service, but also to what extent the (quality) media will be able and willing to distinguish themselves from e.g. social media platforms where considerations of diversity of exposure are typically a less central element of the objective behind the service. Finally, much will depend on the extent to which there will be room for the media to compete on factors such as the diversity of their recommendations. Ultimately, this is also where there could be a role for the regulator. Another recommendation of the Dagstuhl seminar was, therefore, that ‘[r]egulators and industry should collaborate to create an economic environment in which the formalization of diversity does not become a race to the bottom. Online news needs to be (self-)regulated in such a way that the societal responsibility ingrained in traditional journalistic organizations is not pressed out of the market.”

This leads to a final point: the potential role of regulators. In an ideal scenario, in particular, the public service media and quality media organizations themselves will recognize the importance of implementing richer and more diverse metrics in their recommender design. As explained earlier, in reality, media organizations will often experience constraints in funding, in the training of their teams, organizational obstacles, room and budget for innovation, in combination with pressure from commercial players that optimize for short-term metrics, such as search engines and social media platforms. Insofar, there is a role for regulators and policymakers to create the conditions, and incentives for technological innovation in the media sector and for academia-industry collaborations. The fact that many innovation projects in the media are financed by private initiatives, such as the Google News Initiative is a positive example of social responsibility but does not reflect well on the state of public funding for news innovation in many countries. Second, there is also a role for states creating the conditions for fair competition and a flourishing and diverse media landscape, also in the light of the growing importance of, and dependence on some of the leading intermediary platforms – a topic that is subject to extensive discussions e.g. in Brussels right now. Third, there is potentially also a role for more guidance on the question of what media diversity is, and the goals it is supposed to serve, as well as how to implement it into algorithmic design. Here we see a clear role for international think tank and expert groups, ideally also in cooperation with international standard setting authorities, such as Council of Europe, the UNESCO or the OSCE, in other words: with institutions that already have a long tradition of formulating criteria and recommendations to states and private players for the conditions that need to be fulfilled to operationalize diversity, also in algorithmic design.

Bibliography

Page details

Date modified: