Tension, torn, or truly stuck? Understanding tensions in multi-stakeholder recommender development at legacy media companies through paradox theory

1 Introduction

In an era of intensifying platformization, legacy media organizations have increasingly turned to recommender systems (RS) to help audiences navigate their vast catalogs of content (Ranaivoson et al., 2024). In the context of this study, “legacy media” refers specifically to established media organizations that traditionally served as major distributors of print and/or audiovisual content, and that have since become major players in the digital distribution of news and entertainment media.

Both public service media (PSM) and private players in small media markets are confronted with how platform logics are fundamentally reshaping the online audiovisual and news landscape (Iordache, 2024). RS adoption serves as a strategic response to these challenges. However, implementing RS introduces new organizational challenges related to ensuring that the algorithms used are aligned with the organizational values of the companies (D’Arma et al., 2024; Van Es and Poell, 2020).

There is growing awareness that RS shape what users see, read, and watch (Lischka et al., 2024), which has fueled attention to value-sensitive RS design (Starke et al., 2024). This approach goes beyond simply generating accurate recommendations and focuses on developing systems that take into account the values that media organizations aim to uphold. However, operationalizing such organizational values is not straightforward, since, within the context of legacy media, these values are subject to continuous negotiation among a multitude of stakeholders, including technologists, editors, and business managers (Ranaivoson et al., 2024; Smets et al., 2022). In multi-stakeholder settings, “value-sensitive” refers to the values and interests that each stakeholder expects from or imposes on the RS design. Each of these actors brings distinct demands, objectives, and professional norms to the table. Consequently, divergent and often conflicting viewpoints might create persistent tensions that hamper the development process of value-sensitive RS (Mitova et al., 2023).

Previous empirical research has shown that there are several potential conflicts surrounding RS implementation in legacy media organizations, for example, between public mission and market pressures (Møller, 2022), between long-term and short-term goals (Bodó, 2019), and between editorial human control and algorithmic curation (Mitova et al., 2023; Sørensen, 2019). Although these studies point to persistent areas of tension, they remain largely descriptive in their discussion of these frictions. To date, it remains unclear whether certain tensions are temporary trade-offs because of ongoing organizational transformations, either-or dilemmas requiring clear-cut choices, or persistent paradoxes between competing yet interdependent demands. Without additional clarity on their nature, organizations might struggle to effectively respond to the tensions that occur between organizational stakeholders in the design process of value-sensitive RS. The various challenges across different organizational levels that can be attributed to RS development, which Ranaivoson et al. (2024) refer to as the media management dilemma, remain underrepresented in the literature.

This paper addresses this research gap by drawing upon paradox theory as an analytical framework to map the intra-organizational tensions related to value-sensitive RS development, and highlight that media managers might frame tensions differently, resulting in different responses or resolutions. To do this, we first review the existing body of research on RS-related tensions in legacy media organizations, as well as review the literature on paradox theory. We then introduce paradox theory as the primary theoretical framework for identifying and mapping intra-organizational tensions in value- sensitive recommender system design, while drawing on sensemaking theory as a complementary interpretive lens to account for how different stakeholders perceive, frame, and give meaning to these tensions. Next, we outline how we translate paradox theory into a deductive coding framework to analyze 20 in-depth interviews with key representatives from two private news publishers in Belgium and the Netherlands and one public service media organization in Flanders (Belgium).

In doing so, we expand the currently limited applications of paradox theory in media business and management studies. To date, paradox theory has been used to study media innovation in general (Horst and Moisander, 2015), and to understand PSM's tense relationships with social media (Olsen et al., 2025). However, no previous research has used paradox theory to analyze tensions related to a specific innovation project, such as RS development.

This study contributes to the scholarship on value-sensitive RS design by providing a theoretically grounded account to map multi-stakeholder tensions observed in practice. We provide empirical evidence for the conceptual constructs previously suggested by scholars (Berti and E Cunha, 2023; Gaim et al., 2022; Schulte, 2022). Additionally, this paper aligns resolve or management strategies prescribed by paradox theory with the solutions articulated by practitioners. By highlighting the individual sensemaking through which stakeholders interpret, frame, and operationalize competing values, we highlight the underlying paradoxical dynamics at play. This article provides conceptual clarity on intra-organizational tensions in value-sensitive RS design, shows that tensions are connected through upstream and downstream ripple effects across levels, and offers insights into not only, understanding, but consequently also managing them.

2 Literature review

2.1 Value-sensitive multi-stakeholder RS

2.1.1 Value-sensitive RS design

The importance of incorporating values in the RS design has been well-established in the literature on RS, approached through the framework of “value-sensitive design” (Burke et al., 2025; Friedman and Hendry, 2019; Friedman et al., 2002). Furthermore, within the field of human-computer interaction (HCI), the focus on values in system design has been emphasized by the development of value-sensitive design processes (Friedman and Hendry, 2019; Friedman et al., 2002). Particularly in the context of RS adoption by (news) media organizations, there is a strong call for attention to values, given the impact of media on users and society. Both from computational and social sciences, there is increasing research into RS in media environments, aiming to understand how value-sensitive design actually unfolds within media organizations [see, for example, Bauer et al. (2024) and Møller (2023)]. To grasp the day-to-day practices related to RS development in media companies, the growing number of multi-stakeholder studies has provided valuable insights into how differing values might complicate the development process of RS in different media settings (Blassnig et al., 2025; Carillon, 2024; Mitova et al., 2023; Smets et al., 2022).

While the academic multi-stakeholder approach refers to a broad range of actors who have a stake in how RS are designed, including users, policymakers and regulators, our focus is limited to organizational actors within the legacy media organizations. More specifically, we limit our focus to internal stakeholders who are directly involved in, responsible for, or affected by RS development. This narrower scope allows us to zoom in on the intra-organizational tensions that arise when competing professional values, organizational priorities, and design considerations have to be negotiated.

2.1.2 Tensions identified in multi-stakeholder RS research

In the specific case of value-sensitive multi-stakeholder RS design in legacy media companies, we note that organizational tensions often relate to a broader trend of disruption, in which the traditional ways of media production and distribution are challenged by changing practices in the face of digital innovation. For example, Sørensen and Hutchinson (2018) argue that the adoption of RS algorithms first raises the question of how traditional values in media should be understood in the age of online content dissemination. Similarly, Møller (2022) finds that when news organizations wish to implement RS, they struggle with the boundaries of journalism and public service, as the traditional ideological beliefs, norms, and values of journalism are challenged by emerging technologies. In the same vein, Sørensen (2019) also points to tensions between vaguely defined but societally relevant editorial values, such as fairness, transparency, or diversity on the one hand, and “the mathematical logic of recommender systems” on the other (p.1). The reallocation of editorial decision-making from humans to algorithms (Sørensen, 2020) and, with it, the transfer of autonomy (Carillon, 2024) is causing tensions within newsrooms. Multiple studies point to a need to balance editorial independence (Iordache et al., 2025; Smets et al., 2022) and control (Mitova et al., 2023; Sørensen 2019), with automation. Previous work has identified journalistic hesitation (Mitova et al., 2023) towards algorithms, and scholars explain this through a lack of communication or common language (Grün and Neufeld, 2021; Mitova et al., 2023). Although these works call for more collaboration between the different stakeholders within the organizations to overcome these issues, the identified tensions and their resolves are only discussed at a higher level of abstraction (Van Es and Nguyen, 2024). In what follows, we provide an overview of the tensions described in the literature, following the translation process from values to metrics, which, as described by Smets et al. (2022), characterizes the RS development process during which most tensions between organizational stakeholders occur.

First, as RS design in the context of media organizations can be described as value-sensitive, it comes as no surprise that some tensions are linked to those values. When designing RS, conflicts might arise when organizational stakeholders pursue different goals or values (Van Es and Nguyen, 2024). Tensions might also result from different interpretations of values by various actors (Carillon, 2024), or when those values are insufficiently or poorly defined (Sørensen, 2019). This links back to difficulties related to defining and safeguarding traditional editorial values in the process of RS development (Bastian et al., 2021), where tensions can arise when there is no common ground on the role of RS in relation to the editorial decision-making. Linked to this, previous work points to a tension between recommending content based on “what users want” versus “what users need” according to editors (Blassnig et al., 2025; Fields et al., 2018). This highlights the difficult balancing act between, for example, user expectations versus public mission, between diverse recommendations versus “more of the same” (Lassen et al., 2026), or in maximizing reach while maintaining distinctiveness (Sørensen and Hutchinson, 2018).

Second, previous works have identified tensions when translating values into concrete objectives and key performance indicators (KPIs). Sørensen (2020) raises the question of how KPIs exactly evaluate public service, and Grün and Neufeld (2021) similarly emphasizes the difficulty of translating PSM values into concrete KPIs. Linked to this, several authors point to a tension between the pursuit of short-term results and long-term organizational goals. Bodó (2019) considers this a conflict between immediate user data and the public missions. Smets et al. (2022) highlights that this discrepancy between short-term and long-term goals becomes apparent when formulating concrete recommendation tasks that need to correspond to organizational objectives.

Finally, multiple scholars address tensions related to measuring RS performance. Previous works show that the evaluation and development of metrics is a key area where organizational tensions surface. Van Es and Nguyen (2024) demonstrate that recommender success is often assessed through simple engagement metrics, which fail to fully capture editorial values. This critique echoes earlier concerns that organizational values are complex constructs, and tensions often relate to their conceptualization and operationalization (Carillon, 2024; Clark et al., 2023; Sørensen, 2019). Grün and Neufeld (2021) argue that public values are not easily quantifiable and the evaluation thereof is more complex than the commonly used RS metrics. This is why Fields et al. (2018) calls for increased attention to the metrics that RS at PSM are optimized for, since those are crucial for their outputs. In order to avoid conflicts or tensions related to RS evaluations (Iordache et al., 2025; Smets et al., 2022; Van Es and Nguyen, 2024), the values, KPIs, and corresponding metrics should be discussed, in some way, among all relevant organizational stakeholders.

Overall, each of these existing studies consistently points to tensions between organizational stakeholders in the design and implementation of value-sensitive RS in media companies. However, most of these works discuss such frictions at a relatively abstract level (Van Es and Nguyen, 2024), often grouping different forms of tensions under a single umbrella, or treating different forms of tensions interchangeably. Nevertheless, some of the published works referred to above use terminology that alludes to the relevance of paradox theory. For example, Van Es and Nguyen (2024) state that “there are no straightforward approaches to resolving the contradictions and tensions between optimization-oriented personalization through recommender systems and news media's obligations towards the public” (p.4). Similarly, Sørensen (2019) adopts the term contradictions when describing that “recommender systems have been built to serve individual interests by selecting and filtering content, while public service broadcasting traditionally has aimed to reach the whole population with content that editors think is beneficial for citizens. Bringing the two entities sounds like a contradiction in terms”. (p.1). In another paper, Sørensen (2020) notes that “dilemmas are identified where PSB ideals and requirements clash with the industry praxis of recommender systems” and outlines “the paradox of personalization of public service broadcasting” (p.97). The use of terms such as contradiction, dilemma, and paradox draws our attention to paradox theory as a theoretical framework to help gain clarity on the tensions described in previous studies.

2.2 Paradox theory

In the field of strategic management and organizational studies, paradox theory has received significant attention and proven valuable to move beyond a high-level description of various tensions as conflicts or frictions (Schulte, 2022). The systematic literature review by Schad et al. (2016) shows that most academic works, across various fields, treat all tensions as equivalent, and many studies “simplify the complex, often messy phenomena of paradoxes” (p. 3). They propose paradox theory as a meta-theoretical lens that deals with principles of tensions and their management across multiple contexts, theories, and methodological approaches (Schad et al., 2016). More recently, Berti and E Cunha (2023) reiterate the potential of paradox theory as a meta-theoretical perspective for organizational studies (p.879). We draw upon paradox theory as a meta-theory to bridge media management literature with scholarship on RS, which currently mainly focuses on technical design in computer sciences and editorial values in media studies, while paying less attention to the intra-organizational dynamics through which RS are developed. However, Bednarek et al. (2021) warn that such interdisciplinary work might face challenges of superficiality, lack of focus and consilience with theoretical frameworks and concepts are combined without sufficient attention to their theoretical and analytical boundaries. Therefore, in what follows, we first introduce the different types of tensions discussed in paradox theory literature, which serve as theoretical vocabulary for this study. Secondly, we introduce sensemaking theory as complementary analytical lens that allows to understand how different media managers perceive, frame, and give meaning to these types of tensions.

2.2.1 Theoretical spectrum of tensions

Within paradox theory, tensions are defined as conflicting situations in which two or more competing demands are present simultaneously and are perceived as difficult to reconcile (Gaim et al., 2022; Schulte, 2022; Smith and Lewis, 2011). Each of these tensions can occur at multiple levels: at the micro or individual level, at the level of a specific project, at the organizational level, and finally, at the macro level of the industry, including collaborations or governance bodies (Schad et al., 2016; Schulte, 2022). What makes paradox theory an interesting avenue to explore is that the literature also accounts for how tensions are addressed and dealt with. The prescribed responses range from defensive, short-term solutions that prioritize or suppress certain logics to active resolves, which are more reflective approaches that aim to engage competing logics more sustainably, for example, by integrating them or by treating their coexistence as beneficial (Gaim and Wåhlin, 2016; Jarzabkowski et al., 2013). Paradox theory does not view competing interests as problems to be solved per se, but as dynamic and manageable tensions (Gaim et al., 2022). Despite its name, paradox theory moves beyond “the study of proper paradoxes” (Berti and E Cunha, 2023, p. 879).

The theoretical framework aims to study different kinds of tensions that are distinguished based on how competing logics relate to one another (Gaim et al., 2022; Schulte, 2022). Six different types of tensions can be arranged on a spectrum ranging from either-or framing to both-and approaches: contradictions, dilemmas, trade-offs, dialectics, dualities, and finally paradoxes, as presented in Figure 1.

The literature on paradox theory includes each of these types of tensions to understand the interrelationships, what features they have in common, what distinguishes them from each other, and how each of these tensions can be managed. However, multiple scholars have highlighted the difficulties of articulating definitions for each of the different types of tensions within the paradox theory framework (Gaim et al., 2022; Gaim and Wåhlin, 2016; Schad et al., 2016; Schulte, 2022; Smith and Lewis, 2011). Gaim et al. (2022) argue that these concepts, although distinct, are entangled empirical phenomena that can be somewhat ambiguous, recognizing the interpretive nature of tensions, which means that how tensions are defined is constructed through managers' sensemaking.

2.2.2 The link with sensemaking theory

Paradox theory acknowledges that organizational stakeholders attribute meaning to competing values and frame tensions as either resolvable, persistent, or recurring (Schulte, 2022). This suggests that tensions are not objective properties, but are shaped by how stakeholders interpret and frame competing values, which is reflected in the work of Berti and E Cunha (2023), who argue that “actors’ mindsets, cultural frames and practices lead them frame salient paradox as dualities (…)” (p.875), and Fairhurst et al. (2016) who find “actors combine and enmesh tensions” (p.6). The wording used by these scholars suggests that the conceptualization of tensions differs between actors, meaning that what some people might consider a trade-off may be experienced by others as a paradox (Gaim and Wåhlin, 2016; Stoltzfus et al., 2011). If tensions are indeed shaped through interpretation, then sensemaking theory helps us to understand how the ways in which different stakeholders frame tensions can vary based on their background, education, role, etc.

Sensemaking refers to the process through which individuals interpret ambiguous or changing situations when ongoing activities are interrupted (Weick, 1995; Weick et al., 2005). In organizational settings, a change in circumstances and innovation, such as RS adoption, might interrupt the established patterns or workflows and ignite movement over time (Van de Ven and Poole, 1995). In such times of organizational change, stakeholders individually interpret what is at stake for them, which values may be under pressure, and what possible consequences the changes may have. In this study, we consider sensemaking as individual interpretative processes through which different actors involved in RS development perceive organizational dynamics and conceptualize the related tensions. Integrating paradox theory with sensemaking theory emphasizes the roles of managers in dealing with intra-organizational tensions. Depending on how managers interpret and frame competing logics, the resulting tensions will be dealt with differently: conceptualizing a tension as a dilemma or trade-off often invites an either–or solution aimed at achieving a quick fix (Jarzabkowski et al., 2013), whereas viewing a tension as a duality or paradox leads to a both–and orientation in which competing rationales are engaged simultaneously (Schulte, 2022).

2.3 Applicability of paradox theory

2.3.1 Paradox theory to study multi-stakeholder value-sensitive RS design

According to Schulte (2022), paradox theory posits three contextual factors that reinforce competing logics, causing tensions and, in doing so, making them salient. The first factor is change, which is associated with altering dynamics related to the market and technological infrastructure. This factor directly mirrors the implementation of RS at legacy media companies, as these algorithmic tools radically innovate the way media content is distributed online. The second factor is plurality, which refers to the multitude of organizational goals due to a diversity of stakeholders' opinions (Schulte, 2022). Scherer et al. (2013) demonstrate that the presence of multiple actors adds to the complexity of dealing with tensions, and Schad et al. (2016) highlight that “complexity involves a broad array of stakeholders and interest groups holding competing yet equally valid demands” (p.5). This underpins the applicability of paradox theory to the multi-stakeholder setting in which RS development takes place. The final contextual factor that facilitates tensions to surface is scarcity, which refers to limitations on the resources available to the organization (Schulte, 2022). As highlighted by previous work, RS adoption is impeded by media companies facing budget and time investment constraints (Vandenbroucke and Smets, 2025; Verachtert et al., 2023). Each of the three factors is reflected in the specific setting in which this study is situated: RS development is part of the technological changes and innovation ongoing at media organizations, which, as explained above, is characterized by different internal stakeholders, each with different goals competing for the same limited resources.

2.3.2 Conceptualization of paradox theory for RS

Within paradox theory, tensions are defined as conflicting situations in which two or more competing demands are present simultaneously and are perceived as difficult to reconcile (Gaim et al., 2022; Schulte, 2022; Smith and Lewis, 2011). Throughout this research, we broaden the definition from competing demands to competing logics to encompass all rationales, ideas, expectations, and assertions related to the implementation of RS, how the system should work, and how its results should be evaluated. Previous works in organizational studies and management research have highlighted that the literature on paradoxes lacks a uniform conceptual and theoretical foundation (Gaim et al., 2022; Gaim and Wåhlin, 2016; Schad et al., 2016; Schulte, 2022; Smith and Lewis, 2011). To address this, Smith and Lewis (2011) proposed a set of definitions for various constructs to avoid fragmentation of existing research. More recently, Schulte (2022) and Berti and E Cunha (2023) highlight that the current body of work still lacks definitive conceptual boundaries. Therefore, to adopt paradox theory as an analytical framework, we need to clarify the conceptual boundaries of the different types of tensions within the scope of this paper. We do not treat these types of tensions as fixed categories that can be identified independently from actors' interpretations. Rather, we recognize that tensions are “empirical phenomena” whose nature is constructed through the way stakeholders interpret and frame them (Gaim et al., 2022, p. 1984). Our framework (Figure 2) provides an overview of all the concepts we are adopting, and serves as a baseline to map the intra-organizational tensions related to value-sensitive RS design. It includes the six different types of tensions (introduced in Section 2.2.1) that can arise across four distinct levels: the micro-level (individuals), the project level (the RS itself), the organizational level, and the macro-level (industry and governance) (Schad et al., 2016).

Building on previous works (Gaim et al., 2022; Gaim and Wåhlin, 2016; Schulte, 2022; Smith and Lewis, 2011), and following the order presented in Figure 1, we consider the following definitions: We then further distinguish between four types of paradoxes, following the categories proposed by Smith and Lewis (2011) that correspond to the way in which each paradox originates within the organization. The first type of paradox is theparadox of belonging. This tension arises from conflicting identities, roles, or group memberships, such as when individuals or teams need to act autonomously while also belonging to a larger organizational whole, or when multiple organizational roles clash (Jarzabkowski et al., 2013; Schulte, 2022). The second type is theparadox of learning, which surfaces when exploiting existing knowledge, routines, or systems and exploring new, experimental, or radically different approaches (Jarzabkowski et al., 2013; Schulte, 2022). The third type of paradox is theparadox of organizing. This one relates to tensions stemming from conflicting professional values, workflows, or decision rights between departments or disciplines when coordinating between groups (Jarzabkowski et al., 2013). The fourth and final type is the paradox of performing, which occurs when tensions arise related to how work is done, especially the need to learn skills, unlearn old routines, or adapt workflows, often in response to innovation or change (Jarzabkowski et al., 2013). Identifying these tensions is the first step toward understanding how they are managed.

• Contradiction = two opposing logics that cannot exist at the same time, where choosing one necessarily excludes the other. The opposition is framed as mutually exclusive, not something that can be balanced or combined.
• Dilemma = an either–or choice between two alternatives that are equally desirable or equally undesirable, where selecting one option means explicitly sacrificing the other.
• Trade-off = a situation in which gains in one area gradually lead to losses in another, often framed along a continuum. The tension is expressed as a matter of degree (more versus less), not a binary choice.
• Dialectic = an ongoing cycle of opposition and change, where one solution generates a new counter-solution, and progress is framed as emerging through synthesis or transformation over time rather than resolution.
• Duality = two interdependent logics, where neither can be addressed in isolation, and as a result must be managed together.
• Paradox = contradictory yet interrelated logics that coexist over time, are seen as persistent, and are framed as something that must be accepted and managed simultaneously, rather than solved or eliminated.

In paradox theory, for each of the types of tensions, resolve strategies are proposed, distinguishing between defensive and active resolves (Gaim et al., 2022; Gaim and Wåhlin, 2016; Jarzabkowski et al., 2013). When a tension is framed as a dilemma or trade-off, literature suggests it would be resolved through contingency, which involves weighing the pros and cons of either side and subsequently prioritizing one pole over the other (Schulte, 2022). This results in a short-term fix that eliminates or avoids the tension, typically by prioritizing one logic over another. When competing logics are seen as a duality, the defensive resolve proposed involves repeated cycles of compromise, causing the tension to resurface or intensify over time. Such sequences of short-term fixes ultimately result in avicious cycle (Gaim and Wåhlin, 2016; Schulte, 2022). Contrary to defensive resolves, active resolves aim to understand the roots of the tension and overcome it once and for all. Paradox theory highlights deliberate efforts to understand the origin and nature of tensions as the best way to develop more sustainable, long-term ways of addressing them (Jarzabkowski et al., 2013). When tensions are framed as dialectics, Schulte (2022) suggests an integration response by creating new solutions that combine competing logics, rather than choosing between them, often by redesigning practices, roles, or systems. Finally, to deal with paradoxical tensions, literature points to creating a virtuous cycle that deals with tensions by acknowledging and embracing competing logics. These resolves leads to creative outcomes, mutual reinforcement, or sustained improvement over time (Gaim and Wåhlin, 2016).

In summary, paradox theory functions in this study as a theoretically sensitizing framework: it provides the theoretical basis, conceptual delineations, and systematic vocabulary needed to examine the intra-organizational tensions that surface when RS are developed and implemented in legacy media organizations. Its established concepts guide our theory-driven framework and deductive coding process, allowing us to map recurring tensions in a systematic way. At the same time, we do not treat the concepts and boundaries from paradox theory as fixed or exhaustive. Instead, we combine paradox theory with a sensemaking lens to examine how individual media managers interpret, frame, and give meaning to these tensions from their situated background, expertise, and organizational role. This allows us to analyze managers' interpretations while leaving room for empirical accounts that nuance, complicate, or diverge from the conceptual definitions in the paradox literature, including cases where tensions are framed differently or where alternative response strategies are proposed. This position will also be reflected in the methodological design of the study, where our conceptual framework (Figure 2) is translated into a theory-driven codebook which serves “as a means of organizing text for subsequent interpretation” (Fereday and Muir-Cochrane, 2006, p. 83).

3 Material and methods

3.1 Research design

This paper presents a secondary qualitative analysis of interviews originally collected by the authors for two separate research projects, aimed at understanding how different internal stakeholders conceptualize and evaluate recommender systems at legacy media organizations. The interviewees were originally selected purposive sampling based on their involvement in, and specific knowledge of the development and implementation of the RS at their respective media companies. Each of the participants is directly involved in decision-making regarding RS, which provides privileged access to certain information (Van Audenhove and Donders, 2019). The aim of this work is to include well-informed organizational stakeholders who can provide rich insights into the RS development from different perspectives. The original interviews were conducted until saturation in the data was achieved. Although during the first analysis some tensions between stakeholder perspectives surfaced, those tensions were not the primary analytical focus at the time [see Van der Elst et al. (2025) and Vandenbroucke and Smets (2025)]. In this current study, we return to the interview transcripts with a new interpretative lens. We argue that secondary analysis of previously collected data is a valid and useful way to understand the tensions inherent in RS development processes, as illustrated in the work of Bowler et al. (2018), who re-analyze existing transcripts to assess the explanatory power and applicability of a new theoretical framework. Our sample consists of 2 in-depth, semi-structured interviews with internal stakeholders involved in RS development within legacy media organizations. These interviews were conducted between November 2023 and March 2024, at three media organizations in Flanders and the Netherlands, of which two are multi-national, commercial news publishers, and one is a PSM organization. These in-depth interviews lasted on average 68 min. We acknowledge that data collected through individual interviews are retrospective and declarative, meaning that they capture how organizational actors describe and justify RS-related tensions, rather than how these tensions unfold in real-time interaction or how actors behave in practice. Interviewees' accounts may therefore diverge from observed behavior, as respondents may emphasize certain tensions, rationalize past decisions, or frame their own role strategically. However, the aim of our study is not to reconstruct sense making as an interactional process, or observe how tensions are dealt with, but to examine how individual media managers articulate and frame tensions and resolves, from their situated professional roles and experiences.

Therefore, our sample includes participants who range in their role within the organizations. Five stakeholders have a purely technological profile (such as project manager RS, data analyst, head of data), two stakeholders perform a business role (including marketing, or acquisition and conversion), and five stakeholders hold an editorial position (such as editor-in-chief). In addition to actors who fit under one professional logic, the sample also includes eight stakeholders with hybrid roles. An overview of the respondents and their role within the media company can be found in Supplementary Appendix 1.

All quotes that we present are originally formulated in Dutch and are translated by the authors.

3.2 Data analysis

We adopted a qualitative approach to analyze the interview transcripts combining deductive and inductive coding, as outlined by Fereday and Muir-Cochrane (2006). In a first step, we conducted a deductive mapping of the RS-related tensions using a theoretically grounded coding framework derived from prior research (Beighton, 2021; Fereday and Muir-Cochrane, 2006). Our coding framework reflects our conceptualization of paradox theory as presented in Section 2.3.2. In the second step, we analyzed how interviewees frame and interpret these tensions. This step was informed by sense making theory, not to reconstruct how sensemaking emerged as an interactive process during RS development, but to understand how individual organizational actors articulate tensions from their situated professional backgrounds and organizational roles. This enabled us to examine how stakeholders attribute meaning in dealing with competing values and frame potential tensions as resolvable, persistent, or recurring, and whether their framings aligned with or diverged from the paradox-theoretical categories. After mapping the types of tensions, we reflected on what resolve practices the literature prescribes for each of these tensions and to what extent they correspond to the solutions mentioned by our interviewees. Taken together, this approach can be characterized as abductive. While the analysis is primarily structured by the deductive application of a paradox theory as an analytical framework, it also leaves room for inductively identifying variation in how stakeholders themselves framed, evaluated, and responded to RS-related tensions in practice.

The coding was performed by two coders. To support analytical consistency between the two coders, the coding scheme was developed based on definitions from the literature, which were translated into clear, operational coding constructs (Feucht, 2016). These coding constructs enable us to analyze how the theoretical concepts are applied in practice. The complete codebook is included in Supplementary Appendix 2. Both coders first independently coded the same two transcripts using the initial codebook. They then met to compare their coding across both transcripts. Whenever discrepancies arose, these were discussed in detail, and decision rules were added to the coding manual. For example, clarifications were added to distinguish between “duality” and “paradox”, and between the “paradox of learning” and the “paradox of performing”. After updating the coding manual, both coders again coded transcripts independently and met to discuss and compare their findings.

This iterative process of independent coding, comparison, discussion, and refinement of the coding manual was repeated until all 20 transcripts had been analyzed.

4 Findings

To unpack the organizational tensions surrounding value-sensitive RS design, we performed a structured mapping approach grounded in our coding framework. We deductively analyzed all tensions mentioned by respondents during the interviews, linked to the implementation of RS in the organizations. First, we discuss the RS development dynamics in each of the organizations in our sample. Second, we report our mapping of the identified tensions, and we outline the tensions that occur most at the organizational, RS-project, and individual levels. Finally, we compare theoretically prescribed solutions with the resolve practices expressed by the respondents.

4.1 Contextualizing RS development

To understand the tensions identified in the interviews, we first situate the development of value- sensitive RS within the specific organizational contexts of our sample. We recognize that the tensions mapped in this study are unique to the specific situational contexts of each organization and individual sense making expressed in each stakeholder's interview. Nevertheless, our findings show that the same tensions emerge across these organizations when navigating RS development.

4.1.1 Public service media

For the PSM in our sample, RS development is primarily anchored in its centralized online content portal. Since 2016, the PSM's management contract with the Flemish Government mandates them to employ a public service algorithm that helps users discover new, diverse content without neglecting the importance of shared experiences, dubbed a “taste-broadening” algorithm. Its editorial logic is rooted in the public service remit to present audiences with qualitative and distinctive offerings that reach and reflect all of society. Similar to other European public broadcasters recently, the PSM we examine has adopted a digital-first strategy (Van der Elst et al., 2025), which is characterized by the shift from a broadcast-first mindset (grounding their strategy in linear broadcast channels) to a user-first mindset (grounding their strategy in user preferences). In this context, the technical design choices, such as how diversity and/or universality are operationalized, are closely tied to the broader discussion of public service values. Although this paper does not focus on the algorithmic architecture, these design and evaluation choices possibly shape how organizational actors understand the role, purpose, and possible risks of RS.

4.1.2 Commercial news publishers

For the two commercial news organizations in our sample, RS development currently appears to be in a stage between experimentation and fully integrating RS into online news dissemination. Automated recommendations and personalization emerged as exploratory features around 2019 and have since evolved into a core component of their digital distribution strategy. As demonstrated in previous works (Vandenbroucke and Smets (2025); Smets et al., 2022; Mitova et al., 2023; Blassnig et al., 2025), the difficulty lies in aligning the many objectives for RS, such as user engagement, conversion, and subscription retention, into a single design. This requires alignment among a wide range of organizational stakeholders. Here, the editorial logic represents journalistic integrity and traditional journalistic values, which meet market-oriented imperatives and requirements related to advertising revenue and subscription sales.

4.1.3 Unified analysis of tensions related to value-sensitive RS design

We argue that, although PSM and private news organizations differ in their RS development trajectories, both contexts can be analyzed together. Respondents described a similar set of tensions shaping value-sensitive RS design. In both settings, RS development unfolds in a multi-stakeholder environment in which different organizational logics, such as editorial, business, and technological logics, must coexist within a shared digital-first strategy. Both editorial logics seek to reconcile audience preferences with broader societal objectives. PSM operate under a remit to inform, educate and entertain while maintaining the audience reach necessary to fulfill that remit effectively.

Commercial news media, by contrast, are primarily driven by the need to generate revenue, which incentivizes them to cater to audience preferences, while they simultaneously retain democratic and societal responsibilities as journalistic institutions. Although these editorial logics differ in their institutional grounding, they deal with comparable tensions in RS development. RS, and by extension all automation and innovation initiatives, are expected to enhance audience reach and engagement, while upholding editorial judgments about which content is relevant, valuable, or socially important. This shows that tensions related to RS design, implementation, and evaluation arise across public service and private media organizations alike. While technical design choices are a crucial part of RS development, we approach them as organizationally embedded decisions that can make intra-organizational tensions more or less salient. Our interviews suggest that many editorial and business stakeholders are not primarily concerned with the exact technicalities of system design as long as these do not visibly affect their own responsibilities, workflows, editorial practices, or strategic priorities. Therefore, this paper focuses on the different types of tensions that arise between organizational stakeholders when RS are developed, evaluated, and integrated into legacy media organizations.

4.2 Overview of types of tensions

To report our findings, we mirror the two steps from our coding process. First, we outline the different types of intra-organizational tensions surrounding value-sensitive RS design. Second, we examine the resolve practices described by our respondents, comparing how they articulate managing these tensions in practice with the solution strategies prescribed in the literature.

4.2.1 Contradictions and dilemmas

At the start of the “either-or” to “both-and” spectrum are contradictions and dilemmas. Here, the first tensions we found occur at the organizational level and concern the choice between proprietary data infrastructure and external solutions. Respondents from PSM describe the need to choose between the flexibility of their own infrastructure or participating in EU-wide cooperation with a fixed solution, as both options are perceived as mutually exclusive with no middle ground. In contrast, respondents from commercial news organizations described a similar choice between self-built RS models and third-party solutions as a dilemma. In their framing of this choice, these stakeholders recognize both options as equally desirable and capable of temporary coexistence. They explain that third-party systems are particularly useful in the beginning stages of RS adoption, when they are still exploring the possibilities of personalization, and then later they shift to designing the model in-house. This decision relates to the limited financial resources available to media organizations, where investing in an innovative tool offered by a third party is seen as a cheaper way to initiate RS adoption before the budget is freed up to develop the tool in-house.

Additionally, we observed another organizational tension, highlighted by PSM respondents, which involves the struggle between a “broadcast-first” and a “user-first” mindset. Stakeholders explain that these opposing logics require a definitive strategic choice: the organization either maintains a creator- led, linear broadcasting logic or adopts a user-centric logic. The depiction of these two mindsets as a binary choice between mutually exclusive elements, where choosing one pole necessarily entails rejecting the other, leads us to label this as a contradiction. Here, it is important to note that respondents are referring to the underlying logics. Although these mindsets are binary and contradictory, their strategic implications do not necessarily have to be. For example, when opting for a digital-first and user-centric mindset, a paradoxical tension between linear and digital distribution strategies may surface, as both types of distribution have very different characteristics, but both have to be managed simultaneously. This shows that stakeholders adopt a different framing or interpretation of these related tensions, depending on the level at which they are perceived.

4.2.2 Trade-offs

Trade-offs are prevalent throughout our interviews. Both PSM and news publishers express difficulties in dealing with a trade-off of investments. At the organizational level, different innovation projects compete for the same limited budget. The complete budget available can be understood as a fixed whole, and when a larger share is allocated to RS development, less remains available for other projects, and vice versa. At the RS-level, respondents across all three organizations in our sample emphasize that RS adoption is impeded by budget constraints:

“Personalization simply costs a lot, so if you can't prove its success, it will lose trust and the money will go elsewhere” (Tech 5) This illustrates that a lack of proof of the success of RS might lead to a loss of trust and a redirection of budget. The level of time, budget, and attention allocated to the RS shapes perceptions of its strategic importance across stakeholder groups. However, several respondents argued that further technological improvement and optimization may be unnecessary when the current RS is already “good enough” (Business/Editor 1). This ties to whether the RS is treated as a strategic priority or merely a support tool, as one respondent from a news company explains:

“We need to discover how technology can support us, but I think we will always remain critical, precisely because we are news brands, about the fact that it will truly be a support and will never take over what we do”. (Tech/Editor 2) Similarly, PSM stakeholders emphasize that RS will not fully replace manual curation, as Business/Editor 3 explains that currently their homepage curation remains strictly “a human process, we don't leave that to algorithms”. Other respondents, however, explain that rather than one extreme, there is a balancing act between RS and human curation that “can be expressed as 50–50, but the pendulum is constantly swinging” (Tech 6). This framing of curation as an either-or choice between editorial and algorithmic selection implies a trade-off, where at times the balance shifts more toward human editorial curation, and at other times more toward RS, while the total curation remains 100%.

Notably, respondents do not separate this trade-off from the question of positional prominence on the (home) pages. Because page space is limited, choices must continuously be made about what is shown, where it is positioned, and who is in charge of filling these placements. We observed that stakeholders often refer to these positions in a highly structured way by counting them:

“We conducted tests with the 5-6-7-8th position, setting up the recommender there”. (Tech 4) This shows that through assigning numbers, these positions are framed as slots that could be allocated either to editorial or algorithmic curation. This way of talking suggests that homepage curation is understood as a finite and ordered space. Furthermore, we note that different terms are used to describe the most prominent areas of the homepage:

“Personally, I am in favor of personalizing the homepage, but only from a certain position. Everything above the scroll is curated by editors and then below that by the algorithm”. (Business/Editor 1) The framing of “above the scroll” suggests that some stakeholders draw on digital-first imaginaries when making sense of homepage curation. Other stakeholders have stuck with a print-based imaginary of those top positions and label them “above the fold”:

“We never personalize the entire homepage. Purely because the editorial team indicates what is important for you to know, regardless of your interests. So, above the fold, we will always put what you need to know, and below the fold is just what you probably want to know”. (Business 1) We argue that these linguistic choices are not merely incidental but reflect differences in individuals’ sensemaking. This finding echoes the work of Van Es and Poell (2020) on how Dutch PSM actors understand and organize their daily activities in relation to algorithms and data infrastructures.

Regardless of how one defines this high-visibility space, the question of what is displayed is intertwined with who is tasked with filling these positions. Similarly, our respondents often articulate the trade-off between either human or RS-curation in the same breath as the allocation of positions: “It will never, ever be the case that the entire homepage will be based entirely on personalization. What I see happening is that, in the next step, we will say that not the top position, nor the second and third positions, but the fourth or fifth positions will feature articles based on your reading habits”. (Tech 3) Although the trade-off between human and algorithmic curation is most visible at the level of the RS, the way this balance is negotiated depends on the openness of different teams and, more specifically, of individual actors. As one respondent noted:

“Today, it is largely curated by humans. Some desire to have recommendations also appear more on the homepage, but some editorial teams are not open to this yet”. (Tech 2) This suggests that the extent to which algorithmic curation is allowed to expand is not determined solely by product-level decisions, but also by intra-organizational tensions surrounding how teams and individual actors respond to the redistribution of curatorial authority. These findings align with prior multi-stakeholder research, which has shown that tensions arise with the shift in editorial decision-making from humans to algorithms (Sørensen, 2020), leading to conflicts about control over curation (Mitova et al., 2023; Sørensen, 2019), editorial independence (Iordache et al., 2025; Smets et al., 2022), and autonomy (Carillon, 2024).

A related trade-off concerns not who curates, but on what basis decisions about curation are made. Here, tensions arise around the extent to which decision-making should be based on data insights versus editors’ gut feelings. The increasing availability of user and content performance data provides many insights, yet stakeholders also reported frictions concerning the influence on final decision-making. They emphasize that on the flip side of data insights, there are experienced editors who make valuable assessments rooted in years of professional expertise.

“We found ourselves relying heavily on experience and gut instinct, even when it came to personalization and segmentation. And although I had a sense of it, I didn't quite get a grip on the fact that we weren't making enough of an impact with all the personalization options we had. So, our data insights team conducted a deep dive into all the data available at the time, and that gave us some guidance”. (Business/Editor 3)

4.2.3 Dialectics

We identified a dialectic regarding the definition of a “good” recommendation: “It's difficult to discuss the relevance of recommendations, because what we consider important in newsrooms, what people should read, is not always what they want to read. So, is it a good recommendation, once again a Taylor Swift story, or rather yet another article on the situation in Gaza? There are so many articles, but can you imagine that in terms of social relevance, some articles might be rated a little higher than others, but then again, the Taylor Swift stories might be more popular, so which do you want to be recommended?” (Business/Editor 2) Although this respondent is from a news organization, this ongoing tension between, on the one hand, recommending socially relevant content that reflects the media's educational role, and, on the other hand, recommending content that generates more engagement and high click-through rates, thereby tailoring to users' preferences, also emerged at the PSM. We label this as a dialectic where the thesis is content “people should consume” and the antithesis is content “they want to consume”. Rather than framing this as a contradiction, there is an ongoing cycle between editorial logic (the thesis: recommending what people need exposure to) and market logic (the antithesis: recommending what people want exposure to) aimed at the search for a synthesis that combines the educational role while meeting the audience's needs and interests.

4.2.4 Dualities

Following paradox theory, we use the term duality to refer to tensions between two interdependent logics, where neither logic can be addressed in isolation and both therefore need to be managed together. This differs from a paradox, where opposing elements need to coexist over time and therefore require simultaneous, ongoing management. For dualities, the element of interdependence is central to our analytical interpretation, rather than persistent coexistence: the elements involved are mutually influential, so a change in one affects the other, which makes the tension fragile and difficult to manage. When interviewees describe elements as connected, mutually dependent and both necessary while difficult to reconcile, we label the tension as dualistic.

The first duality we observed throughout the interviews concerns the perspective on RS itself: it is simultaneously treated as a technical system that must be developed and maintained as a robust system “in its own right” (Business/Editor 3), and a strategic “means” to achieve broader organizational goals. These two perspectives can seem competing, yet they are interdependent:

RS cannot contribute to organizational objectives without being developed as a reliable, high- performing system. Conversely, the technical development required cannot be justified without proving its contributions to the organization's purpose. This perspective on RS as supporting organizational goals leads directly to the second duality: the interrelatedness of KPIs and metrics.

While stakeholders can prioritize different objectives, often perceived as conflicts, these goals are mutually dependent. Respondents point to how organizational stakeholders tend to prioritize different KPIs, yet these are not isolated from one another. Rather, they are, as one respondent puts it, interrelated: “Do logged-in users lead to more engagement? Does engagement lead to more logged-in use? It all seems to be somewhat related”. (Editor 1) This illustrates how KPIs should not be discussed as separate demands, since in practice, they influence one another. Another example is expressed by PSM stakeholders who suggest that reach is a precondition for retention, and both objectives are, in turn, necessary preconditions to enable users to encounter taste-broadening recommendations. This tension appears at the organizational level rather than solely at the level of the RS-project. We argue that, even in the absence of an RS, tensions would remain between stakeholders prioritizing different KPIs, because the underlying issue is balancing interconnected objectives. In addition to KPIs, we found dualistic tensions also emerge in relation to metrics:

“Ensuring that RS makes people read more articles within a session is currently the most important objective, but we still need to discuss which metric represents this: is it attention time in a session, the number of articles clicked, or a page CTR?” (Tech 2) This reveals a tension between stakeholders who foreground different evaluative metrics, even though these metrics are themselves interrelated and influence one another. Furthermore, tensions arise not only over which metric to use, but what that metric signifies. One respondent explains that click-through rate (CTR) was often put forward as a key metric precisely because of its connection to other metrics and KPIs:

“There are many metrics that contribute to a KPI, and I think CTR is a starting point. While the editorial team places more weight on which content is recommended, the business side will want to know more about how to increase the page CTR, and then you also have the CTR on the recommended items, which gives us insights about how that recommended block performs compared to another block or list”. (Tech/Editor 2) Our findings echo prior multi-stakeholder research by Iordache et al. (2025), Smets et al. (2022), Van Es and Nguyen (2024), who highlight that tensions arise when different stakeholder groups pursue different values, KPIs, and metrics. Our conceptualization of these tensions as dualities reveals that a shared goal, for instance, increasing engagement during a session, may be interpreted differently by different stakeholders. This suggests that although KPIs or metrics are shared, in practice, different stakeholders might interpret them through the lens of their own roles and responsibilities. This reflects the coexistence of multiple, but competing interpretations that must be managed together over time. These findings resonate with prior work showing that when differing interpretations of organizational values are not dealt with, it can result in abandonment of metric development altogether (Carillon, 2024).

4.2.5 Paradoxes

When tensions are described as two or more contradictory yet interrelated elements that coexist over time, and something that must be accepted and managed simultaneously, rather than solved or eliminated, we label them as paradoxes. Following paradox theory, we adopt four categories of paradoxes.

4.2.5.1 Paradox of belonging

This type of paradox arises when tensions are related to the individual identity, role, or group membership of an individual or team. Throughout our study, this type of paradox surfaced, for example, in editors' hesitation to engage in an interview on the topic of RS. Some editors, who were suggested by colleagues as being involved in the RS development, indicated that they felt insufficiently knowledgeable about RS to participate in an interview. This points to a paradox of belonging at the individual level, where professionals who directly work with RS simultaneously feel that they lack the expertise or legitimacy to do so. A second tension that respondents highlighted during interviews concerns editors' fear of being replaced. Previous works have highlighted similar tensions on losing control over curation when RS take over human editorial selection (Mitova et al., 2023; Sørensen, 2019), yet our mapping of these tensions as paradoxes of belonging contributes to research by showing that resistance to RS is rooted in professional identity, legitimacy, and sense of belonging. Respondents link these concerns to skepticism toward personalization and AI in general:

“There is a tendency to quickly form negative perceptions about personalization”. (Tech 1) “People project their concerns and fears onto AI”. (Business/Editor 1) These examples illustrate how paradoxical tensions surrounding belonging may be difficult to identify and manage. For practitioners, this implies that RS adoption requires attention to the anxieties, role insecurities, and perceived threats that may remain hidden. Based on our findings, we advice managers to develop shared dashboards with role-specific explanations, annotated recommendation outputs that explain RS workings, and recurring collaborative meetings in which multiple stakeholders jointly evaluate the role of RS.

4.2.5.2 Paradox of learning

In our analysis, we identify the second type of paradox when tensions arise between exploiting existing knowledge or routines and exploring new, experimental, or radically different approaches. In the case of value-sensitive RS development, we observe paradoxical tensions extend beyond the RS-project level and occur at the organizational level linked to a broader transition toward data-informed media production and distribution. Earlier, we touched upon the trade-off between data-driven and editorial decision-making, and we also observed paradoxes of learning connected to this. Dashboards and performance reports are increasingly available, but editorial teams often lack the literacy needed to interpret them. One respondent explains that:

“There are editorial dashboards available, which we built ourselves, because those reports about the recommender are a bit of a black box for the editorial team. Generating a report is one thing, but if it doesn't provide them with any insights, they are useless”. (Tech 2) Another stakeholder similarly notes that editors are often reluctant to act on these data insights, which harks back to the belonging paradox discussed earlier:

“Editorial teams and, how should I put it, data savviness are not a good combination. This also means that there is a certain bias among editorial teams that you have to take into account, and they should actually be aware of this themselves. Editors are not very enthusiastic about algorithms because they don't feel very confident in that area, which is why they are reluctant to discuss it”. (Tech 5) Identifying these tensions as paradoxes of learning shows that RS adoption is embedded in broader organizational struggles over data literacy and interpretation, reflecting a high-level challenge of digital transformation. For practitioners, this underlines that providing dashboards or data access is insufficient in itself; the challenge lies not merely in providing data, but in rendering data interpretable and actionable across stakeholder groups. Additionally, respondents stress that not all stakeholders are interested in the same numbers, questioning whether they should adapt reports and dashboards to support interpretation for that specific stakeholder group. As one respondent from a news organization explains:

“If your target audience is an editorial team, they have very different needs than if your target audience is a content performance team, or B2B or B2C teams. An editorial team mainly wants to know why a specific article is recommended, and the other isn't, B2B wants information about the advertising revenues”. (Tech 2) This makes organizational learning paradoxical: customized dashboards may improve usability and reduce overload, while shared dashboards may foster common understanding and cross-team learning. Therefore, dealing with this paradox of learning entails finding a good balance between optimizing performance reporting and providing useful insights. By customizing dashboards to show each stakeholder group only what they want to see, one avoids overloading. On the other hand, it costs a lot of money and effort to build all the different dashboards. At the same time, a single shared dashboard may stimulate knowledge sharing across stakeholder groups by showing editors the metrics that, for example, marketing looks at. However, dashboards or reports overflowing with numbers might prevent stakeholders from seeing the forest for the trees. Some respondents highlight this need for data literacy is connected to another specific issue concerning editors’ own news consumption behavior, which might distort their understanding of RS output:

“Editors have a very strong bias because they look at that homepage all day long. They read so many articles, or at least, they don't read them, but they click on so many articles that their recommender naturally starts to behave very strangely, and that results in a discussion”. (Tech 4) This illustrates that editors often consume news in atypical ways compared to ordinary audiences, which can shape and skew the recommendations they receive. Yet, because of limited algorithmic and data literacy, editors lack the reflexive understanding of the impact of their consumption behavior on recommendations:

“The challenge is that journalists and editors are not representative of the general public. They read news headlines differently than you and I. Therefore, their recommendation or personalization lists will always differ from ours”. (Business 1) These findings show that data literacy remains an ongoing learning challenge that affects innovation on the organizational, RS-project, and individual levels. We highlight the paradoxical character of this tension because it requires relying on established expertise of each stakeholder group, while at the same time engaging with new data practices, developing new interpretive skills, and investing in improving data literacy.

4.2.5.3 Paradox of organizing

Paradoxes of organizing concern contradictory yet interdependent demands in coordination across stakeholder groups. In our interviews, we identified three distinct examples of this type of paradox. First, on an RS-project level, respondents point to a tension between treating the RS “as a feature on a website, which was fun to experiment with and to see its potential” versus “a product, a component that we could utilize across all brands” (Tech 2). While the shift from an experimental feature to a strategic product may appear as a choice, we argue it represents a paradox, following the framing of our respondents. Moving “from initial ideas to a product-like vision” (Tech 2) requires exploitation of the current RS in a standardized manner, while simultaneously continuing to explore new possibilities (Vandenbroucke and Smets, 2025). Managing this paradox involves ongoing coordination to ensure that the requirements of ensuring a stable RS that contributes to organizational goals coexist with further RS innovation. This paradox of organizing provides context for two tensions discussed previously: (i) the trade-off of investments, where justification of resource allocation requires return-on-investments, and (ii) the duality of the perspectives on RS, where the RS is viewed both as a robust technical system and a strategic means. A respondent explains that this shift from conceptualizing RS as a feature to a strategic component brings more stakeholders into the process, “ranging from technical colleagues to senior management, you have to get them all on board” (Tech 2).

This leads to a second organizing paradox: the coexistence of divergent organizational logics. Throughout the interviews, respondents repeatedly point to tensions between fundamentally different stakeholder groups that must coexist, for example, between the business logic and editorial logic. We recognize that a multitude of logics is inherent to organizations. However, although divergent logics are constitutive of organizational life, participants framed the presence of competing logics as a source of tensions. They describe how concrete RS design and implementation decisions require them to coordinate across competing, yet interrelated logics, responsibilities, and expectations that need to coexist. For example, for PSM, the editorial logic is rooted in a public service remit to provide a qualitative, distinctive offering that reaches and reflects all of society. In this context, the RS must balance the business-oriented need for engagement with the editorial mission of societal impact. For the commercial news publishers in our sample, the editorial logic centers on journalistic integrity. Here, RS design needs to increase user engagement and subscription retention while simultaneously safeguarding traditional journalistic principles against market-oriented imperatives.

Finally, paradoxes also emerged within stakeholder logics themselves. For example, the editorial logic is not monolithic. Rather, it is comprised of competing priorities that must be managed simultaneously. In the case of PSM, respondents brought up tensions that occur when multiple brand priorities need to be coordinated within a shared digital environment. In a linear reality, each brand would manage its own delineated outlet channel following its own editorial curation, but in a shared online content portal with a single “home”-feed that is not channel-oriented, tensions arise:

“What they do live on air, is their own business. But from the moment that comes to [the portal] we have to ensure that it all comes together. (…) The more you put onto [the portal], the more discussion you get about: I think this is important, but I think that's important”. (Editor 6) Similar intra-logic tensions also surface in commercial legacy news organizations, where editors differ in their openness to RS. Some editors strongly embrace innovation and are eager to experiment with new tools, while others remain skeptical of technologies and prefer to continue their established journalistic practices, and many editors position themselves somewhere in between. The same applies to the business logic, where this stakeholder group is often presented as one unit, but in reality, it is comprised of an advertising-driven logic and a subscription-focused logic that diverge in their goals and even have opposing objectives.

4.2.5.4 Paradox of performing

The final type of paradox describes tensions related to the organizational workflows. We extend the conceptualization of the paradox of performing to move beyond tensions related to the performance of human actors and include the performance of tools. Here, we observed a paradox of performing at the RS-project level. Respondents express expectations of the recommender to perform well to gain the trust of editors and deserve higher positions on the platform. To achieve those results, it requires precisely such visibility to generate the traffic needed to perform well:

“The recommender is currently integrated in very hidden places on the site, low on an article page, far down in a small right-hand column, where not many people see it. Only when the algorithm comes more to the fore and curates parts of a homepage, will we see what effect and impact it could have”. (Tech 2) We label this the prominence paradox, showing how RS must prove their value to gain a more prominent position on the (home) pages, while simultaneously, that very performance is hindered by denying recommendations to be displayed in a prominent position.

4.3 Resolve practices

This section compares the management strategies described by our respondents with the resolves prescribed by paradox theory literature. Following the spectrum of tensions mapped in the previous section, we report the resolving strategies: from defensive resolves to “either-or” choices to active resolves to encourage “both-and” approaches.

4.3.1 Defensive resolves

Defensive resolves include contingencies, which are short-term solutions to tensions by choosing one option over another. Based on the situation, pros and cons will be weighted, and managers will opt to temporarily prioritize one element over the other. For tensions framed as contradictions, such as the choice between proprietary infrastructure and external solutions for PSM, the proposed response is selecting one of the elements that are perceived as mutually exclusive. For dilemmas, such as the transition from third-party to in-house RS faced by commercial publishers, is resolved through the selection of one option, but allows a temporary coexistence of both options.

This resonates with paradox theory, which prescribes contingency resolves to deal with dilemmas and trade-offs. For the trade-off between investments at the organizational level, respondents explain the temporary choice to allocate budget to one innovation project over another: “Ultimately, there is only one thing we can do due to our limited resources, so at some point you have to make choices: are we going this way or that way? And then people usually just look at what delivers the most business value”. (Tech 3) Rather than addressing root causes, such defensive resolves prioritize immediate manageability.

When such short-term solutions are repeated, they might become a vicious cycle of continuous compromising, where choosing one element over another increases pressure on the opposite element(s), causing the tension to resurface or intensify over time. When navigating the trade- off between algorithmic curation and human editorial curation, respondents explain selecting one option over another, and ensure spatial separation:

“The easiest way to keep everyone happy is to have the editorial teams choose which, how, and where the recommenders are placed”. (Tech 6) “The field in which we are allowed to play with the RS is very well defined; it is limited to a few positions at the bottom” (Tech 1) While this proposed solution may temporarily reduce the perceived tensions that arise with the shift in editorial authority and autonomy, it does not address the underlying causes. Instead, it displaces the tension by limiting the visibility of recommended items, which may simultaneously stall further technological development and organizational learning.

In contrast to such a defensive approach where the RS is relegated to fixed, lower positions on the page, stakeholders suggest shifting the framing of curation from an either-or trade-off toward a both-and paradox. This involves a deliberate effort to understand the root causes of the tension and develop more sustainable, long-term ways of addressing it. Rather than viewing manual and algorithmic curation as competitors for finite space, these respondents argue for a collaborative relationship. As one stakeholder from a news organization emphasized: “The one does not exclude the other, we must view it as both-and” (Business/Editor 1). At PSM as well, there appears to be a shift in sensemaking towards RS as an enrichment or augmentation to editorial curation, often referred to in the literature as “algo-torial” practices (Bonini and Gandini, 2019). This reframing constitutes an active resolve, instead of a defensive spatial compromise, stakeholders pursue a coexistence.

4.3.2 Active resolves

Active resolves include deliberate efforts to understand the root causes of tensions and to develop more sustainable, long-term ways of addressing them. The most common example of an active resolve is an integration, which aims to solve tensions by creating new solutions that combine the competing elements or logics, rather than choosing between them. Literature on paradox theory suggests such an effort to develop a new alternative is suitable for dealing with dialectics (Lövståal and Jontoft, 2017).

Throughout our interviews, active resolves emerge as a deliberate effort to synthesize competing logics in the RS design in one integrated solution. According to a respondent of a news organization “a hybrid solution is always the best” (Tech 5). Beyond technical design, integration was achieved through the physical organizational (re)structuring, where literally integrating one team into another aims to facilitate collaboration:

“The closer IT is to the newsrooms, the better. It's quite new that they are embedded in the newsrooms, with data scientists and developers sitting among the editors, creating a kind of natural flow when they work things out together; so if they have a question, they can ask it”. (Tech 5) “The content performance team is now located at the center of our newsroom, it creates one dialogue and people who know each other” (Editor 2) When dealing with paradoxes, literature suggests a virtuous cycle that acknowledges and embraces the competing logics and manages them simultaneously to generate creative outcomes, mutual reinforcement, or sustained improvement over time. In our interviews, this was visible when respondents expressed: “I think you have to look at content and data together” (Editor 5), or “we ensure that our ideas and their expectations align”. (Tech 6). This approach aims to combine both logics without an integrated solution, but rather keep both logics fully intact and evaluate them in a timely manner.

Our findings suggest that the relationship between the type of tension and the chosen resolve is not always linear. While literature suggests that paradoxes require long-term virtuous cycles Gaim and Wåhlin (2016), in one instance, a respondent described a temporary response to address a tension that we classify as paradoxical. When discussing competing demands of the commercial department and the editorial team, he explained “you can dismiss marketing for longer than your editorial team” (Tech 1), which indicates a short-term resolve in which the interests of the editorial team are prioritized over business demands. While this appears to be a defensive move that prioritizes one pole, it might reflect a momentary framing of this tension as a dilemma instead of a paradox: rather than two interrelated logics that must be managed together over time, an immediate choice is made.

This suggests that in this specific situation, a temporary contingency resolve might be the preferred response, even if it does not address the paradox at a deeper level.

5 Discussion

This study set out to map the nature and possible resolutions for commonly occurring tensions encountered when implementing value-sensitive RS in a multi-stakeholder environment at legacy media companies. Our case study at three media organizations in Flanders and the Netherlands demonstrates that tensions, inherent to RS development in a multi-stakeholder setting, are dynamic and dictated by stakeholders' individual sense making. By moving beyond the descriptive mapping of tensions, this section shows the interplay between paradoxes, trade-offs, and other types of intra-organizational tensions at different levels. Additionally, our analysis reveals that the framing of these tensions by stakeholders determines the management interventions.

5.1 How short-term choices might unlock virtuous cycles

At the organizational level, traditional media companies face challenges in their digital transformation due to scarce resources. The RS competes with other innovation projects for limited budget and time. This situation manifests first as a trade-off of investments, which is defined as the necessity of allocating a finite budget to one project at the expense of another. Alongside this, organizations navigate a paradox between technological robustness and their strategic legitimacy.

While the trade-off focuses on the “either-or” of resource distribution, the paradox reflects a “both-and” requirement: the RS cannot contribute to organizational objectives without technical performance, yet its technical development cannot be sustained without proven strategic utility. In the literature on paradoxes, resolving tensions through short-term contingency as a defensive resolve is often characterized as a negative management strategy that ignores the underlying tension (Gaim and Wåhlin, 2016). However, our findings suggest a more nuanced perspective. Figure 3 illustrates how the choice to prioritize investment in the RS primarily improves its technological aspects, such as algorithm development and data processing scalability. We suggest that such short-term allocations may create the conditions under which a virtuous cycle becomes possible.

For this virtuous cycle to emerge, stakeholders must reach agreement on which KPIs and metrics are used to evaluate RS performance consistently across departments. On a technical level, the development of the algorithm, testing its functionality, and data processing all present challenges regarding the speed at which recommendations are generated and the scalability of the model.

Furthermore, the system requires moving beyond hidden positions to generate the traffic necessary to provide evidence of organizational value (see Section 5.3). This process is deeply contingent on editorial trust, ensuring that the technical development is perceived as a means to support organizational objectives. When organizational stakeholders agree on which metrics demonstrate whether RS contributes to shared KPIs and thereby proves its added value at the strategic level, the willingness to further invest in RS development is likely to increase. In this sense, provided that those alignments are met, a one-off choice to invest may soften the organizational paradox, potentially moving toward both technical growth and functional legitimacy, which reinforce one another in a virtuous cycle.

However, our findings show that currently, different stakeholders within each of the legacy media companies in our sample prioritize different KPIs and metrics. Therefore, Figure 3 is best understood not as a fully established causal sequence, but as a conceptual synthesis of the potential long-term dynamic suggested by our findings. We highlight that, provided there is at least some alignment around what counts as successful RS performance, investing in RS may make a virtuous cycle possible when editorial and business actors recognize its contribution to organizational goals, and the system is granted the visibility to test its potential.

5.2 Dealing with the human editor vs. recommender tension

Our analysis shows that tensions can be framed either as trade-offs or as paradoxes, and that this framing can even evolve through managerial sensemaking. We illustrate this with the tension identified between manual editorial and algorithmic curation in Figure 4.

When media managers view this tension as a trade-off, they frame it as an either-or where a specific (home)page position is filled by a human or by an RS. However, other interviewees would see this as a paradox in which one does not exclude the other, but rather combines both curation practices. Following paradox theory, framing this as a trade-off calls for a contingency resolve, where one option is selected over the other. Our interviews show that in practice, an either-or selection often results in the RS being relegated to hidden positions (see Section 5.3). This defensive resolve might trigger a vicious cycle rooted in the paradox of belonging: editors perceive the RS as a threat to their professional identity and fear being replaced by automation. This loss of control over curation may fuel reluctance among editors toward RS, which Mitova et al. (2023) describe as journalistic hesitation. This, in turn, hinders the improvement of data literacy, which is associated with a lack of understanding of the output of RS, which can further reinforce the sense of loss of control. This dynamic leads to a vicious cycle.

However, when framing this tension as a paradox, stakeholders will aim to achieve integration through enriched or augmented curation. This both-and approach to curation seeks to build a hybrid practice that combines tailored recommendations via RS with editorial judgments. Our findings suggest that, in practice, this collaborative view on curation is still in its early stages. We observe that it is currently a deliberate management intervention aimed at evolving toward algo-torial practices (Bonini and Gandini, 2019). In this both-and perspective, editors are less likely to experience lost authority but rather assume a more augmented role in which they collaborate with the system. This, in turn, improves editors' understanding of RS outputs, fostering a virtuous cycle of increased literacy and trust. Our study, thus, demonstrates that the framing of tensions is crucial to understand how they are managed.

In addition to promoting shifting sense making of the same tension from trade-off to paradox, managers can adopt other interventions to mitigate the vicious cycle or reinforce the virtuous cycle. Several respondents point out that atypical consumption patterns among editors lead to skewed recommendations, which editors then misinterpret as an RS flaw. To resolve this, respondents explain they use dummy-proof presentations and videos to demystify the black box of the algorithm. These training sessions are most effective when they are repeated consistently and tailored to the specific professional jargon of the target audience (Vandenbroucke and Smets, 2025). Another intervention observed in the organizations was the physical co-location of editors and data scientists in the same office space. This spatial integration is intended to foster daily collaboration, helping to transform the RS from a threatening external tool into an integrated component of the editorial workflows.

5.3 Ripple effect of tensions across levels

By analyzing the tensions inherent in value-sensitive RS development, we find that these conflicts are not isolated but are deeply intertwined, creating a cross-level ripple effect. Tensions at the organizational level (e.g., the trade-off of investments) dictate the conditions for tensions at the RS-project level (e.g., system visibility), which in turn shape individual-level experiences (e.g., professional belonging and trust).

To illustrate this interdependency, we anchor our discussion in the prominence paradox. As identified by multiple respondents, for RS to be granted a high-traffic position, it must first perform well and prove its value; however, to generate the interaction data required to optimize performance and demonstrate results, the RS must be sufficiently prominently displayed. Our findings suggest that this paradox is not merely a standalone operational hurdle, but rather links to other upstream and downstream tensions.

We argue that an upstream tension impacting the prominence paradox is the difficulty in aligning metrics and KPIs and agreeing on what is understood as “good RS performance”. When an RS is evaluated through simplistic or siloed metrics that fail to capture nuanced editorial values [see Carillon (2024); Van Es and Nguyen (2024)], the system's “success” lacks organizational legitimacy.

If the metrics used to justify the initial investment trade-off (Section 5.1) do not resonate with editorial values, the RS will not demonstrate its value enough to move from hidden to prominent positions. This brings us to a downstream tension linked to the prominence paradox: the level of editorial trust. At the individual level, a paradox of belonging surfaces where editors are affiliated with the RS project yet lack the expertise and literacy to interpret its outputs, leading to a fear of losing curatorial control or being replaced. This is further amplified by the paradox of learning at the organizational level, where the knowledge gap between technical output and editorial literacy remains unaddressed.

Without a clear, value-aligned understanding of why the RS suggests certain content, editors fall into journalistic hesitation. Editorial trust in RS acts as a catalyst, where lack of trust leads to denying the system more prominent visibility, and consequently, the RS remains stuck in a low-visibility placement where it can neither improve nor prove its worth.

The prominence paradox illustrates one such chain, but similar cross-level ripple effects are also visible in tensions around metrics, dashboard interpretation, and intra-editorial (dis)agreement over the potential of RS. For example, our findings suggest that reach, retention, and taste-broadening are best understood as a duality. Reach can create the conditions for retention, retention can support repeated exposure to diverse content possibly unlocking taste-broadening, which depends on users encountering and engaging with the RS in the first place. The tension therefore lies not in the need to pursue all three objectives simultaneously at all times, as would be the case in a paradox, but in the fact that prioritizing or changing one objective affects the other. When this interdependence is overlooked, tensions at the organizational level may cascade downward: prioritizing only reach, for instance, may encourage RS-project decisions aimed at maximizing short-term exposure, whereas prioritizing only retention may, in turn, lead to a narrower focus on serving already engaged users with more of the same content, at the expense of discoverability or taste-broadening. Rather than manage these KPIs as interdependent elements, the prioritization of one objective will impact the RS design and evaluation choices, which in turn shape how individual actors interpret the RS output.

Here, the ripple effect unfolds from a lack of strategic alignment of KPIs into RS design choices and influences stakeholders' individual sensemaking and view on RS. Taken together, these findings suggest that organizational tensions related to value-sensitive RS development do not occur in isolation, but unfold in an ecosystem of downstream and upstream ripple effects. Our analysis shows that higher-level strategic choices, often made in response to dilemmas or trade-offs, shape the everyday practices, working conditions, and professional realities of individual actors. At the same time, the success of such choices is also conditioned by upstream effects. When tensions at the individual level, such as low trust in the recommender, limited algorithmic literacy, or reluctance toward data-driven approaches, remain unaddressed, they may become bottlenecks that constrain the implementation and improvement of the RS at broader organizational and project levels.

6 Conclusion

This paper examined how tensions surrounding value-sensitive recommender system (RS) development in legacy media organizations are understood and managed in an intra-organizational multi-stakeholder setting. Drawing on paradox theory, we perform a deductive analysis of 20 interviews with key representatives from one PSM organization and two commercial news publishers. This study identified an intertwined set of intra-organizational tensions that emerge across organizational, RS-project, and individual levels and are interpreted differently by different stakeholders. Rather than treating these frictions under a single umbrella, this paper showed that they can be more precisely understood as contradictions, dilemmas, trade-offs, dialectics, dualities, and paradoxes, with each conceptualization having different implications for how they are perceived and addressed in practice. The analysis further demonstrated that these tensions are not isolated, but connected through cross-level ripple effects, where strategic choices at one level shape tensions and practices at other levels. Our findings suggest that rather than treating tensions as isolated disagreements, media managers could benefit from more anticipatory decision-making which involves considering how tensions are interconnected, which actors are likely to be affected, which values may come under pressure, and which additional tensions may be triggered by a possible decisions, before interventions are made. Such an approach would not eliminate tensions, but it could support more informed organizational negotiation and collaboration around value-sensitive RS development.

Our findings confirm that paradox theory offers a useful framework for mapping tensions in the case of value-sensitive RS design, while highlighting stakeholder sensemaking as a crucial element in understanding how the same tension might be framed differently by different actors, and therefore invites different responses. In this way, the paper contributes to the scholarship on value-sensitive RS design by providing conceptual clarity on intra-organizational tensions and how they are interrelated. A second contribution lies in aligning theoretically prescribed resolve strategies in paradox theory with the management practices articulated by practitioners, thereby offering a more grounded understanding of how such tensions are handled in practice. More broadly, the present study demonstrates that understanding RS in real-world legacy media companies requires attention not only to algorithms and outputs, but also to the organizational dynamics through which values are conceptualized, translated, and operationalized into technical systems. By showing that tensions differ in nature, interact across levels, and are managed in ways that do not always align with theoretical expectations, this paper underscores the importance of treating value-sensitive RS development as an ongoing process of organizational negotiation rather than as a one-off design challenge.

A limitation of this study is that, due to its scope, it does not include audiences as relevant stakeholders. Although many tensions identified in the interviews concern what ought to be recommended to users, those users themselves are rarely directly involved in such negotiations. In a media landscape characterized by technological disruption and growing concerns about the trustworthiness and legitimacy of institutions such as PSM and news media, we acknowledge it may be particularly valuable to consider audience value beyond established metrics (Barr, 2025).

This challenge has already been identified in scholarship on PSM (Urbániková and Smejkal, 2025; Vanhaeght and Donders, 2021), in journalism studies (Swart et al., 2022), and, increasingly, in value-sensitive multi-stakeholder RS research (Ekstrand et al., 2025). Future research could therefore extend this work by examining how audiences themselves perceive, evaluate, and potentially shape the tensions surrounding value-sensitive RS development in real-world settings. Additionally, our empirical scope is limited to three legacy media organizations in Belgium (Flanders) and the Netherlands. While this narrow focus allowed us to examine intra-organizational tensions in depth, it also limits the generalizability of our findings. Future research could replicate our approach in other media organizations, media markets, or national contexts to examine whether similar tensions emerge and how they are shaped by different organizational structures, or media systems. Our study focuses specifically on tensions related to RS development. Although this provides a clear analytical focus, it also means that broader organizational tensions around digital transformation such as AI adoption or data analytics fall outside the scope of this paper. Future work could therefore apply a similar mapping approach to map the tensions linked to a wider range of technological and organizational changes in media organizations. Finally, our study is based on a secondary analysis of retrospective interviews conducted between November 2023 and March 2024. A longitudinal research design would be valuable to examine how tensions and their management evolve over time, including how certain tensions become more salient, fade into the background, or are possibly framed differently as RS development moves from experimentation to implementation and integration.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Generative AI tools ChatGPT Pro and Gemini were used for proofreading of the final manuscript. The data collection, data analysis, and all ideas and insights in the paper are original from the authors and were conducted without the use of generative AI.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fcomm.2026.1840975/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Bruno Dumas, University of Namur, Belgium

Reviewed by: Diaa Ahmedien, Capital University (formerly Helwan), Egypt

Cassandre Burnier, Université libre de Bruxelles, Belgium