Personalizing the war: Perspectives for the adoption of news recommendation algorithms in the media coverage of the conflict in Eastern Ukraine

Implementation of system-driven news personalization

In this article, we focus on system-driven personalization, the origins of which can be traced back to the rise of recommendation systems (RSs) in the 1990s. Ricci et al. (2015: 1) define RSs as software tools that ‘provide suggestions for items that are most likely of interest to a particular user’. Inspired by the growing need for (online) information filtering, RSs quickly became a powerful means of helping users to find relevant information in multiple domains. Following the increasing adoption of RSs by search engines (Hannak et al., 2013) and social media (McGee, 2013) in the 2000s, legacy media organizations also turned towards content personalization to counter cannibalization of their audience and increase ‘stickiness’ (Thurman and Schifferes, 2012) of their websites.

The logic behind news recommendation algorithms is similar to the one used by RSs in other domains such as music or cinema. The increased speed of content production creates conditions under which news readers are not able to consume all available stories. To address the potential information overload, news outlets deploy algorithmic systems which filter news content and order it according to user preferences (Morales et al., 2012). To identify user preferences, personalization systems collect data about different forms of interactions with news content, such as clicks or time spent on the web page and then make predictions based on specific algorithmic designs (Möller et al., 2018).

The accurate estimation of user preferences is a major criterion of personalization effectiveness because wrong predictions leave users unsatisfied and can decrease their trust in the system (Swearingen and Sinha, 2002). Bozdag (2013) identifies several categories of algorithms used to predict user preferences, varying from popularity-based algorithms (ordering news depending on how often they are read) to content-based ones (ordering news depending on how similar they are) and collaborative filtering-based ones (ordering news depending on how similar are users reading them). Content-based algorithms are the most common approach for news personalization, followed by hybrid approaches combining content-based and collaborative filtering (Karimi et al., 2018).

While ways to implement news personalization vary significantly between media outlets, Weber and Kosterich (2018) describe a usual scenario of system–user interactions in the context of news. After entering the system (e.g. by registering on the news website), users are usually asked to provide starting information (e.g. by stating the country of origin) or cross-link their personal data (e.g. by authorizing via a Google account). The system then uses this information to produce first suggestions: for instance, a popularity-based algorithm can recommend stories which currently have the largest number of views, whereas a collaborative-filtering one can suggest a story clicked by users with similar profiles (e.g. of the same age or from the same region). Then, the system collects explicit (e.g. the user’s evaluation of recommendations) or implicit feedback (e.g. time spent on pages recommended by the system) and re-trains the preference model for each specific user.

While the logic behind personalization is relatively simple, the development of an effective news personalization system involves multiple challenges. For instance, there is little information about user preferences before the user actively starts engaging with the system (the so-called ‘cold start’ problem; see Lam et al., 2008). Another challenge is the limited lifespan of news stories that distinguishes news recommendations from recommendations for other domains (e.g. music or movies), where products and preferences tend to be more persistent (Morales et al., 2012). Additionally, the high societal relevance of the news media also raises multiple normative requirements for the design of RSs, in particular when dealing with highly polarizing news content.

System-driven news personalization and media ethics

The importance of addressing the use of news personalization for conflict coverage is related to ethical concerns about the growing deployment of computational technologies by journalistic institutions (Helberger, 2015). Many of these concerns align with general considerations about the use of big data such as acquiring permissions for data use and the proper informing of users (Richards and King, 2014). Other concerns, however, are particularly relevant for news media and deal with the impact of algorithmic systems on transparency of news production and journalistic credibility (Diakopoulos and Koliska, 2016).

The deployment of system-driven personalization raises a separate set of ethical challenges. Individualized news feeds can help citizens to make better decisions and promote media diversity (Helberger et al., 2018), but can also undermine readers’ autonomy and modify their information diets. Furthermore, the obscurity of news recommendation algorithms can erode the audience’s trust (Diakopoulos and Koliska, 2016) and facilitate the use of personalization for manipulating public opinion. The current rise of so-called ‘fake news’ (Lazer et al., 2018) and computational propaganda (Wooley and Howard, 2017) is a telling example of the negative impact that manipulative use of content distribution systems has on the public sphere.

So far, the debate about the impact of news personalization on the public sphere develops primarily in the context of Western democracies. However, such a narrow focus limits the applicability of existing research for other regions, where personalization is increasingly adopted by legacy media. The current lack of studies on news personalization in non-Western contexts leaves behind a number of important questions such as to what degree existing assessments of personalization are specific to liberal media systems and how individualized news feeds can influence the public sphere in countries characterized by limited press freedom.

Ethical concerns about the role of personalization in partially free and non-free media systems are particularly significant in the case of reporting formats dealing with polarizing topics such as armed conflicts. The use of individualized news feeds can enhance media’s capabilities for promoting reconciliation and contribute to peaceful conflict transformation by offering users a more balanced view (Bastian et al., 2019). However, it can also facilitate selective conflict coverage that impedes de-escalating behaviour via the personalized distribution of one-dimensional views on the conflict. In the latter case, personalization can also lead to increased societal polarization and reinforce stigmatization of specific individuals and groups.

Characteristics of the Ukrainian and Russian media systems

Both in Ukraine and Russia, media landscapes are characterized by close ties between the political and media systems, but the nature of these ties is different. In Russia, a ‘paternalistic tradition’ (Vartanova, 2019) has shaped the nature of the media system and led to the introduction of state subsidies and policies intended to guarantee the dominance of federal media. Together with growing authoritarian tendencies in the political sphere, it led to an extremely high politicization and the lack of political pluralism in mass media, as the majority of legacy media organizations became aligned with the state (Vartanova, 2019).

In Ukraine, the media market is dominated by oligarchs, who are often directly or indirectly involved in the political sector (Panchenko and Rybak, 2019). The relationship between oligarch-controlled media and the state in Ukraine is less straightforward than in Russia because Ukrainian oligarchs often use their media to criticize the Ukrainian government. In order to counteract the oligarchic influence, Ukraine has recently introduced laws to improve media ownership transparency; however, the effects of these measures remain limited (RwB, 2019a).

The intense politicization of legacy media leads to significant limitations of press freedom in Ukraine and Russia. In the press freedom index (RwB, 2019b), Ukraine is ranked 102nd out of 180 countries, whereas Russia is ranked 149th. Journalists in both countries experience obstacles in their daily routines, in particular in Russia where violent attacks against media practitioners occur frequently and the legal framework undermines independent journalism (RwB, 2019c). While the state of press freedom in Ukraine has improved after the ousting of President Yanukovych in 2014, the beginning of the conflict in Eastern Ukraine made both Ukrainian and Russian authorities toughen their control on media organizations as a response to the ongoing crisis (Freedom House, 2017a, 2017b).

Despite limited press freedom, legacy media in both countries rapidly adapt to technological innovations. While broadcasting remains the major source of information in the two countries, an increasing number of Ukrainian and Russian citizens use online media to consume news (Deloitte, 2017; RwB, 2016). The digitization of media spheres extends possibilities for independent news outlets (e.g. Hromads’ke in Ukraine, and Medusa ¹ in Russia), but also creates additional opportunities for disseminating fake and propagandist content. These possibilities for media instrumentalization are also amplified by the bilingualism of a large part of the Ukrainian audience. While the use of Russian does not automatically mean an eagerness to consume Russian propaganda, it increases the probability of exposure to Russian media products, which are often instrumentalized by the Russian authorities as part of the information warfare (Makarets, 2017).

Historical background of the conflict in Eastern Ukraine

The conflict in Eastern Ukraine started in the spring of 2014, following the ousting of the Ukrainian President, Viktor Yanukovych, in the course of Euromaidan protests. The crisis of legitimacy of the new pro-Western government (Sakwa, 2014: 129–131) was amplified by the fear of restrictions of Russophone Ukrainians’ rights and worsened by the prominent role of Ukrainian nationalist groups during Euromaidan (Zhukov, 2016). These concerns became a starting point for a series of protests against the post-Euromaidan government, in particular in the Eastern and Southern regions of Ukraine, characterized by a large concentration of ethnic Russians and a high level of support for Yanukovych.

The post-Euromaidan tensions in Ukraine were used by the Kremlin to establish control over the Crimean peninsula and hold a referendum, following which Crimea was annexed by Russia. The annexation prompted a new wave of protests in mainland Ukraine, which culminated in the takeover of government buildings in Donetsk and Luhansk by pro-Russian activists. These actions were followed by the announcement of the formation of the Donetsk and Luhansk People’s Republics, known respectively as DNR and LNR (Makhortykh and Sydorova, 2017). The increasingly violent nature of the protests was amplified by the appearance of armed groups of Russian paramilitaries (e.g. the group led by Igor ‘Strelkov’ Girkin which captured Sloviansk). The growing level of violence forced the Ukrainian government to announce the beginning of the anti-terrorist operation in Eastern Ukraine.

The conflict in the region escalated in May 2014, following the presidential elections in Ukraine and the self-determination referendums in DNR and LNR. The intensification of the confrontation between the Ukrainian army/pro-Ukrainian paramilitaries and pro-Russian insurgents led to multiple casualties among civilians, including the passengers of flight MH17 brought down on 17 July, and the fast deterioration of the humanitarian situation in the region. Despite the intense resistance from the insurgent side, which presumably received military support from Russia (Czuperski et al., 2015: 5), pro-Ukrainian forces managed to recapture the majority of territory held by LNR and DNR until August 2014.

The major change in the course of the conflict happened in the second half of August when, during violent clashes around Ilovaisk, insurgent forces started their counteroffensive. The counteroffensive which, according to the Ukrainian side, was supported by regular Russian units (Czuperski et al., 2015) resulted in heavy losses among the Ukrainian troops, which were forced to retreat. The subsequent advancement of insurgent forces ended with the signing of the first Minsk agreement on 5 September which established a ceasefire between Ukraine and DNR/LNR.

The ceasefire held till January 2015, when a new round of fighting started, culminating in brutal clashes at Donetsk airport and around Debaltseve. The advancement of DNR and LNR forces, which captured Donetsk airport in January and eventually forced the Ukrainian army to retreat from Debaltseve in February, resulted in the signing of the second Minsk agreement. Since then, the conflict primarily transitioned to the trench warfare phase, characterized by less intensity of fighting with a few occasional eruptions of violence such as the ones in Mar’inka in June 2015 or Avdiivka in January 2017. ³

Conflict coverage in Ukrainian and Russian media

From its beginning in 2014, the conflict in Eastern Ukraine has been covered extensively by legacy media in Ukraine and Russia, the two main sides involved in the conflict. Existing studies (Boyd-Barrett, 2017; Fengler et al., 2018) looking at the coverage of the conflict in legacy media indicate significant differences in the way it is presented in Ukrainian/Western and Russian outlets. In the latter case, the predominant narrative revolves around the protection of the ethnic Russian population from ‘the worst of fascist cruelties’ (Boyd-Barrett, 2017), whereas the majority of Ukrainian and Western media interpret the conflict as an instance of Russian aggression.

Despite different interpretations of the war in Donbas in the two countries, structural similarities between the Ukrainian and Russian media systems translate into similar features of the media coverage of the conflict. As in the case of other recent conflicts, media coverage of the war in Eastern Ukraine is characterized by more fragile state control over journalistic narratives compared to the pre-digital time (Pantti, 2016: xvii). In both Ukraine and Russia, digital technology helps journalists to challenge official interpretations of the conflict and provide a more comprehensive coverage of the events. ⁴ Similarly, the use of digital media opens additional possibilities for contextualizing the conflict and circumventing censorship filters (Makhortykh and Sydorova, 2017).

The potential for more constructive conflict coverage, however, is far from being realized. In their comparative study on the media coverage of the conflict in Eastern Ukraine in Ukraine, Russia, Poland and Sweden, Nygren et al. (2018) note that Ukrainian and Russian media tend to present the war in Donbas in rather simplistic terms. The prevalence of the ‘Us–Them’ paradigm is attributed both to official information policies in relation to the war and struggles of individual journalists who try to balance their obligation to be impartial with feelings of patriotism. Under these circumstances, legacy media often adopt the language of state authorities and avoid giving voice to representatives of the opposing side. This self-censorship also leads to the avoidance of news stories which could damage the reputation of the side to which journalists belong (Nygren et al., 2018).

Another important aspect of the media coverage of the war in Donbas are the increasing news dissemination capabilities of Ukrainian and Russian legacy media. Digital news distribution makes the conflict more visible and allows regional media organizations (e.g. Russian NewsFront or Ukrainian StopFake) to share their interpretations with a global audience. Potentially, it also enables a more nuanced understanding of the conflict, as regional media are able to communicate views of different parties to the international community.

At the same time, increased dissemination capabilities also facilitate the distribution of fake and propagandist content. In their study on fake news related to the conflict and disseminated via Russian TV and social media, Khaldarova and Pantti (2017) demonstrate that Russian legacy media frequently employ false stories (e.g. a story about a child crucified by Ukrainian soldiers in Sloviansk) to stigmatize the Ukrainian side. Similarly, in a study of fake news on Ukrainian legacy media, Kitsa (2017) notes that, despite Russian media being more active in distributing fake news, Ukrainian news outlets also occasionally distribute disinformation. ⁵ These observations align with the findings from a study on computational propaganda in the post-Euromaidan period in Ukraine by Zhdanova and Orlova (2017), who argue that both Ukrainian and Russian actors instrumentalize the mass media to manipulate public opinion and facilitate state propaganda efforts.

The media coverage of the conflict in Eastern Ukraine is also characterized by an extensive range of reporting formats. Similar to foreign journalists, who, as Pantti (2019) shows in a study on individual reporting on the conflict in Eastern Ukraine, often use their social media accounts to provide personal commentary in relation to the war, a number of Ukrainian (e.g. Serhii Loiko) and Russian (e.g. Dmitry Steshin) reporters use social media to enhance conflict coverage. A number of legacy media also initiated special multimedia projects devoted to the conflict in Eastern Ukraine such as video series (TSN, 2018) and photo stories (Turchenkova, 2015). This extension of conflict coverage beyond hard news format opens new venues for war reporting which can promote a more nuanced view on the conflict.

This constructive potential, however, is often undermined by the affective language used by Ukrainian and Russian media in relation to the opposing side (Nygren et al., 2018). Highly emotional terms such as ‘terrorists’ and ‘criminals’ are often supplemented with narrative tropes evoking memories of earlier conflicts such as WWII. These parallels are more blatant in the case of Russian media, which often label the Ukrainian side as Nazi successors (Nygren et al., 2018; Sazonov et al., 2016). The Ukrainian media adopt a more sophisticated approach by emphasizing the continuity between WWII Ukrainian fighters and the contemporary Ukrainian army. ⁶

Political system

A number of studies (Hallin and Mancini, 2004; Hanitzsch et al., 2011) demonstrate the impact of political systems on how news is produced and disseminated. The majority of existing studies focus on personalization in Western democracies, where the adoption of algorithmic systems is primarily determined by the media market. In the case of defective democracies, however, the deployment of innovations is dependent on political actors, who have a disproportionate influence on the media sector (Bastian, 2019). In Ukraine and Russia, these actors often treat media organizations as part of their power structures (RwB, 2016: 14), which leads to a slower adoption of technological innovations unless these innovations can be easily instrumentalized for promoting certain agendas.

The nature of the political system also has implications for the design of news personalization. Gillespie (2014) argues that algorithmic systems are not neutral and political valences are inscribed into their designs. Helberger (2019) develops this argument by demonstrating how different models of democracy imply distinct designs of news personalization. In the liberal democracy model, for instance, the best-fit news personalization is particularly relevant in the light of its focus on citizens’ autonomy and sovereignty. By contrast, in the deliberative model, the integration of diversity metrics in the personalization design is essential because this model encourages exposure to different viewpoints. The same logic is applicable for defective democracies, where dominant political values (e.g. paternalism in Russia) influence the functions of legacy media and designs of algorithmic systems used by them.

Legal framework

Regulatory frameworks are important elements of media ecosystems which determine possibilities and limitations for adopting innovations. Amongst other things, these frameworks define norms of data protection and consumer rights that are highly relevant for personalization deployment. Unlike many Western democracies, where regulatory frameworks usually account for the latest technological developments, Ukraine and Russia are slower in recognizing the legal consequences of the datafication of the media sector (Kovalov, 2016; Kukushkina et al., 2017). The processing of personal data remains a rather vague area in post-socialist countries’ legislation but, in both countries, data processing in the context of journalistic activities is partially exempted from existing regulations, thus facilitating the deployment of news personalization.

Legal regulations also influence the selection of news content available for being recommended (e.g. by prohibiting stories detrimental to the state’s image), thus impacting personalization on the content supply level. Both in Ukraine and Russia, topics related to the conflict in Eastern Ukraine are seen as particularly sensitive ones, thus stories produced by the opponent’s side often fall into the category of legally prohibited content (Freedom House, 2017a, 2017b). These regulatory frameworks put additional limitations on the adoption of personalization, in particular in the case of legacy media.

In addition to the selection of available content, the legal framework influences how personalization systems are designed. The growing regulation of algorithmic systems has direct implications for these systems’ design, in particular in the European Union where the GPDR promotes more human interpretability for algorithmic systems (Goodman and Flaxman, 2017). Similarly, the anti-discrimination legislation drives the ongoing debate on integrating fairness into the design of algorithmic systems to prevent their negative social impacts (Lepri et al., 2018). The slow adaptation of post-socialist legal frameworks to datafication so far limits the influence of this factor; a major exception is media regulation in Russia, where increasing attention is devoted to the integration of control mechanisms into the design of automated distribution systems (Matsakis, 2017).

Press freedom

The state of press freedom influences both the range of content available for personalization and the probability of stakeholders instrumentalizing it for their own purposes (e.g. to boost their image through individualized news distribution). In countries with limited press freedom, the deployment of algorithmic personalization depends not only on its potential for engaging the audience, but also its possible effects on state control over the mass media. These concerns are less relevant for Ukraine with its more pluralistic media sphere than for Russia, where authorities exercise significant control over legacy media (Vartanova, 2019). Under these conditions, Russian media organizations can be less inclined to adopt personalization to avoid limiting state control over what citizens read.

The effects of varying degrees of press freedom on the design of personalization can be compared to the ones related to legal frameworks. In the case of systems with a high degree of press freedom, this factor is presumably not going to have a substantial impact on personalization solutions. However, in non-free systems or those that are only partially free, it can lead to a different set of design requirements influenced by local actors’ inclination to instrumentalize these systems. An example of such requirements is the increasing push of Russian authorities for tighter control over the public sphere, which translated into additional design requirements for news aggregators (Matsakis, 2017).

Journalistic values

The production of news is shaped by values that underlie journalists’ work. Determined by their training and the country’s journalistic traditions, these values include both media organizations’ principles and individual journalists’ beliefs. Journalistic values differ between Eastern and Western Europe (Głowacki et al., 2014), in particular in relation to the high degree of journalism’s instrumentalization and its profound connections to the political sphere in the former region. Additionally, both in Ukraine and Russia, journalistic education is slow in adapting to new digital environments (RwB, 2016; Vartanova and Lukina, 2017). Together with a limited awareness of possibilities/threats related to datafication, the weak power distance between journalists and stakeholders (Hanitzsch et al., 2011: 281) can lead to a more uncritical perception of personalization in post-socialist countries.

The importance of the relationship between values and algorithmic design is increasingly recognized in academic scholarship (Vallor, 2016). Not only do developers’ personal values affect their decision-making processes, but also the resulting systems often promote or undermine specific values. The discussion so far is focused primarily on general moral and ethical values but the same consideration applies to journalistic values such as objectivity. Under these conditions, the interpretation of specific values in the journalistic community can have a significant effect on personalization system designs.

Awareness of algorithmization and datafication

The awareness of possibilities and threats of personalization determines the willingness of legacy media to incorporate them into their working routines. Particularly in Russia with its larger and more competitive media market, legacy media increasingly acknowledge commercial benefits of personalization, while devoting limited attention towards its societal implications (Makhortykh and Wijermars, 2019). Seen as an integral means for dealing with the information overload, personalization is referenced as a major innovation in the media sector in Russia (PWC, 2017) and an influential characteristic of the media market in Ukraine (Ishenko, 2017).

Journalists’ awareness of the current state of AI-driven systems is also important for the design of news personalization. Such awareness is essential for understanding which technical solutions can be applicable for specific tasks, and positioning these solutions within the larger structure of the news industry. The limited presence of the filter bubble imaginary in Ukraine and Russia (Makhortykh and Wijermars, 2019) results in the absence of many Western normative concerns and leads to local media’s focus on personalization designs that maximize user engagement for monetization purposes. Comparing the two countries, Russia is a more favourable environment for personalization because of a higher awareness of commercial benefits of datafication and a larger and more innovation-responsive media sector.

Promoting diversity, understanding and journalistic values through personalization

We suggest that a main benefit of personalized conflict coverage is the possibility to enhance information diversity at the user level (Bastian et al., 2019). By tracking the history of user exposure to conflict-related news, personalization can expose users to various viewpoints on the conflict. Individualized news feeds can burst so-called ‘filter bubbles’ by recommending stories that communicate various viewpoints on the conflict or highlight the role of underrepresented actors. By doing so, personalization can mitigate the negative impact of journalistic self-censorship, which influenced conflict coverage both in Ukrainian and Russian media (Nygren et al., 2018). While individual news stories will still be influenced by the attitudes of their authors, the provision of a comparative perspective can decelerate the formation of one-sided views on the conflict.

The realization of such a system will require the use of diversity metrics, such as the ones optimizing for long-tail diversity or serendipity in news recommendations (Vargas and Castells, 2011). Their exact implementation depends on which aspects of conflict coverage are to be diversified. Source diversity is the easiest one: by identifying news sources (information which usually is part of stories’ metadata) to which users are exposed, it is possible to diversify their information diets by recommending stories coming from different sources (e.g. the ones rarely visited by the users) (Bastian et al., 2019). Viewpoint diversity can be implemented in a similar way, in particular considering that post-socialist legacy media often demonstrate distinct editorial lines in reporting on the conflict in Eastern Ukraine.

In cases when a news source can be associated with different interpretations of the conflict, the viewpoint expressed in a specific piece can be detected via lexical analysis. Conflict reporting is often characterized by the recurring use of emotionally charged terms that are indicative of specific conflict parties’ views (e.g. ‘karateli’ is a term usually used by pro-Russian reporters). Finally, actor diversity can be implemented via named entity recognition based on an online referential system (e.g. Wikipedia). By automatically recognizing actors’ names in the text and then identifying their role (e.g. politician or activist), it is possible to extend the range of visible actors beyond political or military elites.

These personalization techniques can counter a one-sided representation of the conflict in Eastern Ukraine (e.g. the Russian aggression vs Ukrainian Nazi schemata, see Khaldarova and Pantti, 2017). They can also expose similarities in the use of affective language between different conflict parties, thus countering the manipulative use of conflict coverage and, potentially, increasing the self-awareness of media practitioners. By doing so, news personalization can also strengthen journalistic values (e.g. objectivity and impartiality) which are hard to sustain at the time of conflict, when journalists often struggle both with official and self-censorship.

Two major obstacles for the practical realization of such diversity-sensitive personalization systems are the limited incentive for Ukrainian and Russian legacy media to implement then and potential resistance from the audience. A limited incentive for the diversification of conflict coverage is related to the fact that post-socialist legacy media often focus exclusively on either pro-Ukrainian or pro-Russian viewpoints and, while doing so, tend to advance stakeholders’ interests. While this tendency makes it easier to use personalization for more independent outlets (e.g. news aggregators), it is still potentially possible for regional media, especially as they already often provide users with a personalized selection of partner-provided news/advertisements. By adding a section of news stories from moderate ideological ‘counter-partners’, these media can increase user engagement by confronting them with ‘“the different”, purposefully or via serendipitous encounters’ (Helberger, 2015: 1328).

Another obstacle concerns the eagerness of the audience to consume attitude-incongruent news. Existing studies demonstrate that direct exposure to attitude-incongruent stories does not automatically lead to their acceptance by the users (Bail et al., 2018). Two points, however, are important: firstly, while not every news reader is interested in a middle ground, online news consumption is still dominated by mainstream and not polarized media outlets (Flaxman et al., 2016). Secondly, instead of confronting users with attitude-incongruent stories, personalization can highlight stories with a high degree of consensus (e.g. the ‘purple feed’ combining views from different sides of the ideological spectrum, see Babaei et al., 2018). The latter option can also be strengthened through self-actualization mechanisms encouraging user interactions with individualized news feeds (Sullivan et al., 2019).

Threatening social cohesion, journalistic professionalism and autonomy through personalization

Personalized conflict coverage also poses a number of threats, including the possibility of undermining journalistic values, enabling masked censorship and increasing societal radicalization. By ignoring values such as impartiality or objectivity, news recommendation algorithms can undermine journalists’ working routines and subvert their societal functions. Similarly, by ignoring or distorting journalistic standards (e.g. by recommending stories from unreliable sources), personalization systems can aggravate negative features of media coverage of sensitive topics such as armed conflicts.

Such threats to journalistic professionalism also include the dissemination of propaganda and disinformation. Individualized news feeds can advance the objectives of actors who have access to the recommendation mechanism by propagating a pre-selected set of interpretations. The use of personalization for purposeful interference in the public sphere opens new options for instrumentalizing conflict coverage. By integrating propagandist stories into users’ news feeds, personalization algorithms can make them more visible and convincing (Bastian et al., 2019).

Complementing the effects of promoting propaganda, personalization can also facilitate the filtering of news items which contradict agendas of specific stakeholders/authorities. By pushing undesired stories out of the public agenda, personalization can enable masked censorship, which is less obtrusive than direct prohibition of certain views (on censorship in Ukraine and Russia, see Freedom House 2017a, 2017b). Under the pretext of satisfying users’ information needs, such individualized censorship can make the audience used to news filtering. In this way, personalization can strengthen stakeholders’ control over journalistic narratives and, together with journalistic self-censorship, limit the possibilities for constructive conflict coverage.

Furthermore, exposure to attitude-incongruent news can exacerbate polarization instead of countering it (Bail et al., 2018). In this way, the use of personalization for diversifying conflict coverage can lead to further radicalization, in particular considering the polarizing nature of conflict-related news. The threat of personalization-driven polarization is particularly pronounced in the case of traditional approaches to conflict coverage which put a strong emphasis on visible effects of violence (Galtung, 2002). By exposing users to violence-focused opponents’ narratives, individualized news feeds can facilitate dehumanization of the opposing side instead of countering it.

The above-mentioned threats are particularly pronounced in the case of simple value-agnostic personalization systems, which are strongly influenced by popularity metrics such as views/clicks. Such systems are frequently employed by Ukrainian and Russian legacy media, and offer limited personalization per se. Instead, they primarily rely on the idea of promoting the most popular content independently of its relationship with journalistic values. These systems are particularly susceptible to third-party manipulations, where stories can be promoted or demoted using views/clicks generated via automated agents. To a certain degree, the same concerns also hold true for collaborative filtering-based algorithms, which can be subverted using fake user profiles that promote specific news items (Burke et al., 2015).