Epistemic welfare and public service media's algorithmic recommender systems: A theoretical framework,operationalization and relevance for governance

Abstract

This contribution introduces the comprehensive framework of epistemic welfare to discuss how public service media (PSM) can engage with algorithmic recommender systems in a manner in keeping with PSM's foundational principles. We contextualize PSM algorithmic recommenders in their tradition of content curation and discuss the challenges PSM face in implementing these systems. We introduce epistemic welfare, a framework based in social epistemology and welfare studies, defined as concerned with creating and maintaining conditions and capabilities for epistemic agency of citizens in the public sphere. We discuss the epistemic standards of reliability, power, fecundity, speed, and efficiency and illustrate the framework's operationalization for the design and implementation of recommenders and its relevance for governance by and of PSM's algorithms. Ensuring that algorithmic recommender systems fit epistemic welfare, we argue, allows PSM to help tackle the epistemic disruptions in the digitalized public sphere.

Public service media in an algorithmic media ecology

Starting from the concept of epistemic welfare and focusing on algorithmic recommender systems, this contribution presents a comprehensive framework and its operationalization to guide public service media's (PSM) engagement with these (and other) artificial intelligence (AI) based technologies in ways that respect PSM's foundational principles. We argue that the ideal and reality of PSM provide opportunities for PSM organizations to contribute to solving what some identify as the epistemic crisis of the digitalized public sphere (Coeckelbergh, 2023; Dahlgren, 2018).

The possibilities, realities and threats of digitalization have taken a prominent position in communication research following the ongoing transformation processes identified as a “trinity of datafication, algorithmization and platformization” (Latzer, 2022: 331). Essentially, algorithms are sets of human-coded or machine-learned instructions that automate selection processes and assign relevance to data to accomplish these tasks (Latzer, 2022). In the field of media, research particularly focuses on the role and impact of recommender systems’ algorithms that suggest content to users based on past behavior, preferences, patterns and content characteristics (Mitova et al., 2023). Acknowledged for helping users find their way in a media landscape characterized by abundance, algorithmic recommender systems, especially those found on platforms, have been criticized for their potential to diminish diversity, accountability, privacy and to contribute to spreading mis- and disinformation, bias and injustice, among others. Crucially, the impact extends beyond information into knowledge of all kinds: what we know, how and why we know what we know. Neuberger et al. (2023: 181) elaborate on how these developments change the “knowledge order,” that is, “processes of generation, verification, distribution, and appropriation of knowledge in public discourse.” They describe the development from a linear knowledge order of which mass media, including PSM, were a crucial part, to digital media's circular knowledge order, seen to flatten epistemic hierarchies and disrupt knowledge practices. While they consider both the opportunities and threats of these epistemic disruptions, for Dahlgren (2018: 23f.) the disruptions constitute an epistemic crisis that undermines trust in the “traditional foundations of knowledge” and destabilizes “the grounds for establishing and legitimizing ‘truth’ [ushering in] an uncertain epistemic future.” While a measure of epistemic anxiety serves as a reminder that some facts require an explanation (Cabrera, 2021), systematic uncertainty regarding what you know, and the how and why of that knowledge, can result in epistemic distress, as observed in the current wave of public distrust toward knowledge generating institutions like education, government and media (Flew, 2021), exacerbated by structural factors.

Like commercial media and platforms, PSM organizations operate in this changing context. They implement algorithm-based systems in their operations to varying degrees, ranging from content production, including live television (Okopnyi et al., 2023), and news collection and composition (Sørensen and Hutchinson, 2018), to recommender systems for personalized content distribution (Álvarez et al., 2020; Hildén, 2022; Perrota, 2023; Van den Bulck and Moe, 2018), the focus of this contribution. In implementing algorithmic recommender systems, both for on-demand services and on (news) websites, PSM organizations find themselves suspended between staying relevant in a crowded and interactive media ecology, providing distinctive content to users-as-citizens and avoiding negative epistemic consequences of recommender systems, all while safeguarding their unique position as trusted organizations (Horowitz et al., 2022; Perrota, 2023; Sørensen and Van den Bulck, 2020). Indeed, as these systems become integrated in PSM operations, they shape the conditions under which trust in PSM is (re)produced, not least because trust is grounded in perceptions of competence, integrity and benevolence (Mayer et al., 1995) and is reinforced through institutional structures, repeated interactions, and normative alignment of its basic characteristics (Zucker, 1986; see Ali et al., 2025).

To meaningfully discuss how PSM can engage with algorithmic recommender systems while living up to its foundational principles, we build a conceptual and actionable framework that combines communication studies with insights from social epistemology and welfare studies. Answering calls to (re-)theorize PSM algorithms (e.g. Iordachi et al., 2025), we develop theoretically and operationalize for PSM algorithmic recommender systems the notion of epistemic welfare which revolves around conditions and capabilities aimed at ensuring citizens’ agency in the public sphere (Hyzen et al., 2025).

After this introduction, we situate the use of recommender systems in PSM's tradition of content curation, based on the foundational principles of universality and distinctiveness. Next, we discuss how PSM organizations have been implementing algorithmic recommender systems and the related problems specific to PSM as discussed in existing literature. Subsequently, we introduce and explain epistemic welfare as a novel framework to think about recommender systems fit for PSM. We elaborate on epistemic standards and illustrate how these can guide PSM organizations’ implementation and evaluation of recommender systems and show how the framework can guide the governance of PSM's recommenders both for system design (governance by algorithms) and for policymaking (governance of algorithms). As such, we argue that this framework ensures algorithmic recommender systems that best serve PSM's normative goals, allowing PSM to help address the crisis rather than contribute to the epistemic disruptions in the digitalized public sphere.

From airwaves to algorithms: Curation and recommendation in PSM

PSM have always been involved in curation, carefully deciding what content to offer and how to offer it. These decisions have been consistently determined by preconceived ideas and ideals reflecting PSM's foundational principles (Van den Bulck, 2001), more recently identified as PSM values (Martin and Lowe, 2014; Moe and Van den Bulck, 2014; Puppis and Ali, 2023). Developed in the early 20th century under a specific set of political, technological and social-cultural conditions (Scannell and Cardiff, 1991), PSM principles have been subject to much debate. However, the PSM principles remain firmly based on two pillars: universality, both as universal access and universal appeal, and distinctiveness aimed at enhancing social, political and cultural citizenship through quality services and output (Born and Prosser, 2001; Van den Bulck and Moe, 2018). PSM content offerings aim to reflect these foundational principles.

Universality of appeal means that PSM provide a range of programs that inform, inspire and entertain, accommodating diverse interests of young and old, higher and less educated, across a broad and increasingly multicultural and multilingual community. The underlying ideal is that a well-functioning democracy rests on an informed citizenry, best achieved through the simultaneous dissemination of shared messages to all citizens (Born and Prosser, 2001; Van den Bulck, 2001). Furthermore, universal appeal is considered to contribute to the nation as imagined community with shared cultural background and identity (Anderson, 1991). This has played out differently across national contexts and time periods, shaped by structural factors like organizational size and capacity, political system and regulation, and by tensions between unifying national narratives and recognizing socioeconomic diversity, minority cultures and multilingual realities (e.g. Lowe and Savage, 2020). Regardless, PSM content performed a key role in Neuberger et al.’s (2023) linear knowledge order, generating, verifying and distributing political, cultural and social knowledge.

PSM principles have guided not just what kind of content is offered but how it is offered. Universality of access ensures that PSM content reaches every citizen the organization is expected to address (Van den Bulck, 2001). PSM in the monopoly linear broadcasting era was mostly limited to one or more generalist channels providing one-size fits all programming. This was achieved through careful scheduling: a paradigmatic selection and syntagmatic combination of programs that created a flow (Williams, 1974), encouraging people to watch/listen to a range of content, even programs that may not appeal at first glance but that PSM organizations, inspired by their principles, deemed important (Van den Bulck, 2009). Over time, this was affected by PSM organizations’ efforts to stay relevant as their monopoly was replaced by a proliferation of commercial channels, the introduction of thematic channels, streaming services and platforms. PSM adjusted in various ways, mostly by continuing linear generalist channels, for now still PSM's core business, complemented with various linear and nonlinear services (Moe, 2008).

Throughout, PSM organizations have tried to match their principles with audiences’ personal preferences. Originally, PSM's cultural-educational logic meant offering audiences what they need, not necessarily what they want (Van den Bulck, 2001), reducing audiences’ options to a decision to listen/watch or not. The monopoly period more easily allowed for this as there were few alternatives. The post-monopoly, multimedia context dominated by private audiovisual media and, more recently, streaming services and digital platforms centered discussions around PSM's distinctiveness, mostly understood as being distinguishable from commercial media. The notion of distinctiveness is contested (D’Arma, 2018), yet as foundational principle and ambition, PSM “should be distinct by virtue of the functions it performs and the value it brings to society” (Jakubowicz, 2010: 13). Much of the discussion regarding distinctiveness has focused on PSM content (Hendrickx et al., 2019) and on the extent and scope of online PSM services (Bardoel and Lowe, 2008). We argue that distinctiveness should extend to PSM content distribution and, thus, to their recommendation systems.

Datafication, algorithmization and platformization affect PSM's efforts to combine distinctiveness with meeting audiences “where they are,” that is, adapt content curation and distribution to individual users’ preferences. Algorithmic recommender systems allow for advanced personalization of content distribution and services yet force PSM organizations to carefully consider how this can be accomplished while safeguarding universality and distinctiveness. We now turn to the particular challenges these developments pose for PSM's foundational principles.

The challenges for PSM of implementing recommender systems

While private media and platforms are mostly guided by their commercial interest when investing in and profiting from personalization tools and algorithmic fine-tuning (Winseck, 2020), PSM organizations must weigh pros and cons against their foundational principles when incorporating algorithmic recommender systems into their operations (Fields et al., 2018). Concerns regarding adverse effects of such systems, including epistemic disruptions caused by audiences confronted with reduced diversity of content and views, among others, are particularly salient for PSM organizations (Hildén, 2022). PSM have been urged (e.g. Helberger, 2015) to take a proactive approach in implementing recommender systems to help address these concerns and to counter epistemic disruptions by setting relevant standards. PSM organizations, on their end, have been cautious in adopting recommender systems. Research into PSM organizations’ early approaches (Van den Bulck and Moe, 2018) showed various strategies, from embracing algorithmic developments to retaining “old style” linear/analog applications. Importantly, these variations related to contextual factors, including political or financial restraints, rather than differing foundational values.

This initial hesitation has noticeably shifted, pushed by PSM organizations’ need to remain present in their media landscapes. By 2022, 16 out of 56 European PSM organizations had recommender elements or an active recommender system in their on-demand platforms and, since then, several small (e.g. VRT, YLE) and large (e.g. BBC, ZDF) PSM organizations are exploring recommender systems for on demand services and news distribution (Álvarez et al., 2020; Fields et al., 2018; Grün and Neufeld, 2022; Hildén, 2022; Mitova et al., 2023). Some of these systems’ opportunities and challenges are unique to PSM as public service providers (Horowitz et al., 2022; Sørensen et al., 2020), as we discuss below.

Conflicts between personalization and PSM's foundational principles

Personalization through recommender systems can clash with PSM's foundational principles of universality and distinctiveness. First, universality is about making sure that relevant content reaches the entire population. Recommender systems’ individualized approach (Sørensen, 2020a) is problematic when personalized PSM offers something for everyone but not (ever) the same for everyone, potentially contributing to polarization and fragmentation. So, PSM organizations should aim for recommender systems that retain a shared discursive space, central to shared knowledge and community building.

Second, recommender systems can hamper societally relevant content reaching PSM audiences. Changes in viewer habits and competition with commercial platforms may push PSM to follow a neoliberal logic by prioritizing audience preferences over encouraging meaningful cultural relationships (van Es, 2017). If PSM organizations implement recommender systems the same way as for-profit media and platforms, personalization can result in audiences being approached merely as consumers rather than citizens, receiving personalized (and highly enjoyable) yet societally irrelevant content. Finding a balance between reach and societal relevance is crucial. Furthermore, algorithm-driven curation can undermine PSM's significance as inclusive content providers given that algorithmic predictions often fail to serve smaller, for example, minority language, audiences.

Third, and related, recommender systems can undermine PSM's distinctiveness. PSM's programming decisions, that is, distinctive content selections, are not necessarily affected by how this content is (algorithmically) distributed. However, distinctiveness is achieved also through how selected content is offered, in the analog era through scheduling. This carefully curated menu is now replaced by algorithmic recommendations, stripping away the cultural meanings that result from flow. Moreover, as personalization prioritizes content that fits the algorithmic data analysis of audience preferences rather than societal needs, certain distinct content can be marginalized, serving as a disincentive for PSM to invest in such content altogether.

For PSM to meaningfully implement recommender systems it must opt for systems that ensure the preservation of the foundational principles, identified as “public service algorithms” (Michalis, 2022; van Es, 2017). Some argue that, since media organizations already harvest extensive user data to analyze and anticipate users’ needs, preferences and usage history, they could use profiling and targeting with a commitment to social good (Helberger, 2015) or, as we argue, to epistemic good.

Operationalizing public service principles for recommender system design

PSM organizations are well aware of these conflicts and have been addressing them through algorithmic recommender designs that variably focus on diversity, serendipity, coverage, novelty, fairness, transparency, explainability, accountability, sustainability, trustworthiness and/or editorial control (Fields et al., 2018; Grün and Neufeld, 2022; Panteli et al., 2019; Sørensen, 2019). The endeavor remains challenging, as it includes attention to issues (e.g. privacy) that previously were not explicitly part of PSM considerations and as it requires operationalizing abstract principles into measurable metrics. Challenges include balancing popularity metrics with distinctiveness metrics, ensuring exposure diversity across the range of PSM offerings through algorithmic recommendations, and providing transparency of the logic underlying recommendations (Sørensen and Hutchinson, 2018). However, PSM organizations as yet have no clear conception of how to encode goals and principles into their recommendation algorithms (Piscopo et al., 2024). The complexity of ensuring PSM principles also necessitates a large number of, possibly conflicting, algorithmic metrics to be operationalized, while variations in PSM across organizations and countries hamper collaboration among PSM organizations (Sørensen and Hutchinson, 2018).

A further challenge for PSM organizations is the dependence on commercial players like software companies, third-party providers, and platforms. For instance, Sørensen et al. (2020) explored how European public and private media websites employ third party services to track and analyze user behavior and found that PSM organizations, especially those carrying advertising, closely resemble private media's use of such services. This leads to a further operational and cultural challenge: recommender systems are managed by data scientists while expertise in public service principles resides with editorial staff, necessitating a reconciliation of different perspectives (Piscopo et al., 2024).

To safeguard longstanding principles of universality and distinctiveness and incorporate emerging issues like privacy, while fulfilling organizational goals and embracing the potential of recommender systems, we argue that PSM require an overarching principled and actionable framework. This allows them to move beyond compiling lists of values for translation into algorithms, to infuse PSM principles with concrete meaning and to design measures that properly evaluate whether PSM recommender systems promote these principles.

Incorporating audience agency

PSM distinctiveness extends to considering the audience as citizens rather than “mere” consumers. For commercial media and platforms, personalization mostly serves as a tool to keep audiences connected to harvest their data and expose them to commercial communication for profit maximization (Mansell and Steinmueller, 2020). Their recommender systems offer users personalized content based on previous use, similar users’ behavior, explicit user preferences, content popularity and/or content-related characteristics (Mitova et al., 2023), “often without the user's deliberate choice, input, knowledge or consent” (Zuiderveen Borgesius et al., 2016: 3). As such, recommender systems provide users with surprisingly little agency, variably referred to as user autonomy, sovereignty or control, to determine their choice of media diet, thus limiting their epistemic agency, that is, users’ ability to control the formation and revision of what constitutes their knowledge beliefs and interests (Coeckelbergh, 2023).

PSM organizations, in contrast, are expected to serve their audiences as citizens by being distinct from commercial players but also by allowing user control through recommender systems, strengthening citizens’ agency (Sørensen and Hutchinson, 2018). Here, PSM organizations face a conundrum as encouraging user choice can undermine a shared discursive space and facilitate selective exposure instead of diversity (van Es, 2017), exacerbating rather than tackling epistemic disruptions. A recommender system that nudges users to consume a more diverse content menu could be effective, as research suggests users often choose from the first few recommendations (Garu, 2018), but such strategy resembles traditional PSM's paternalistic cultural-education logic (Sørensen and Hutchinson, 2018), undermining users’ autonomy to explore a variety of content and form their own opinions, detrimental to users’ empowerment and agency.

Currently, some PSM recommender systems offer limited user agency, often basic interactions like BBC+ on demand's functionality to like/dislike content. Some PSM news recommender systems provide more advanced features, like YLE's allowing users to search specific topics and rate their importance on a five-point scale (Hildén, 2022). Other options remain unexplored. In making choices, we argue that PSM must ensure not just user agency but users’ epistemic agency. PSM can contribute to empowering citizens to obtain and expand knowledge.

Below we propose epistemic welfare as a normative framework that provides a systematic and holistic approach to PSM organizations’ design and implementation of recommender systems, to guide policymakers and regulatory authorities’ oversight, and to empower audiences. Our framework builds on scholarship on media pluralism and diversity (Napoli, 2001). While existing models emphasize source, content, and exposure diversity, we focus on the conditions under which such diversity becomes epistemically meaningful for users-as-citizens. Following Vrijenhoek et al. (2021), we recognize the importance of embedding exposure diversity into recommender system design and propose epistemic standards as criteria that shift the normative goal from “exposure” to “epistemic engagement.” These standards also address challenges faced by PSM in multicultural societies, where universalist commitments need to be reconciled with differentiated epistemic needs and expectations across diverse publics.

PSM and epistemic welfare

Epistemic welfare: A framework

To create a framework that guides PSM organizations in their engagement with recommender systems, we start from the notion of epistemic welfare, defined as “creating and maintaining the conditions and capabilities for epistemic agency of citizens in the public sphere” (Hyzen et al., forthcoming). We follow Kaun et al. (2023: 878) who consider welfare provision “fundamentally about how societies are organized and what shared values prevail to guide this provision, but also what lives people are able to live.” Different interpretations hereof lead to different welfare regimes, yet in advanced capitalist economies, provisions center around “assistance and support to those citizens who suffer from specific needs and risks characteristic of the market society” (Dencik and Kaun, 2020: 1). The hyper-commercial logic underlying the developments that have led to the epistemic crisis, reducing all human activity to extractable value in the form of commodified data, warrants efforts from PSM to contribute to conditions and capabilities that epistemically empower citizens. This welfare perspective goes beyond epistemic rights (Michalis and D’Arma, 2024) to encompass a broad set of conditions and capabilities.

Conditions for epistemic agency involve the contextual-structural factors that enable or hinder knowledge acquisition and dissemination. They include (digital) infrastructural arrangements, unrestricted and affordable access to knowledge generating systems and organizations, a well-functioning information and communication environment, and policy frameworks that support this (Puppis et al., 2024), next to supportive social and cultural environments (Cetina, 2007). Individuals’ access to knowledge generation, verification and distribution, opportunities for knowledge appropriation, and potential to engage in and shape public discourse fluctuate according to these conditions, affecting their spectrum of empowerment and disempowerment. As we elaborate below, for recommender systems, the ideal and reality of PSM make these organizations well-situated to be an integral part of the conditions that ensure epistemic welfare, following a social responsibility conception of PSM.

Capabilities for epistemic agency denote individuals’ abilities, that is, skills, competencies and critical understanding required to access, appropriate and disperse knowledge (Nussbaum, 2006). They are distinguished by various internal attributes of the individual and represent the abilities that enable people to navigate, engage with, and benefit from knowledge generating systems. Skills like media literacy and digital literacy turn availability (conditions) into access while scientific literacy, critical thinking and problem-solving turn access into knowledge acquisition (Carey, 1986). Communication skills, in turn, ensure an individual participates in knowledge production and sharing (Fricker, 2015).

Conditions and capabilities are both essential to ensure that individuals can exercise their epistemic agency. Without the right conditions, individuals may not have the opportunity to develop and utilize their epistemic agency. Likewise, without the necessary capabilities, individuals may not be able to take full advantage of the affordances available to them. Finally, individuals may choose not to exercise their epistemic agency to pursue epistemically value states.

Epistemic welfare's aim is epistemic agency of citizens in the public sphere, that is, empowering citizens to reach—what social epistemology calls—epistemically valuable states that encompasses (i) having true beliefs, (ii) avoiding errors, (iii) having justified beliefs, (iv) having rational beliefs (or partial beliefs), and (v) having knowledge (Goldman et al., 2011). In our case: grounded in universality and distinctiveness, PSM organizations contribute to citizens’ epistemically valuable states by offering content that helps citizens form true and justified beliefs, through factual information and accurate representations of their cultural context, social world, and personal identities. Through documentary, storytelling, analysis, and diverse representations, PSM can support true and rational belief formation, strengthen understanding, reflection, and a sense of one's place in the world. From a social epistemology perspective, this constitutes epistemically valuable content, that is, content that contributes to citizens reaching epistemically valuable states. Crucially, PSM must not only curate such content but distribute it, including through recommender systems, in a way that contributes to citizens’ knowledge and justified beliefs. Epistemic welfare strives to strengthen citizens’ epistemic agency, creating an environment that allows citizens to pursue epistemically valuable states at their own choice and pace, rather than curbed by profit-driven personalization of commercial media and platforms or by paternalistic perspectives of the old cultural-educational PSM logic.

Social epistemology, like Dahlgren (2018), recognizes that “the availability or lack of certain technologies changes effective standards of knowledge” (Godler et al., 2019: 222), making social epistemology a good inroad to understand (solutions to) the role of recommender systems in epistemic disruptions. For epistemically valuable states to be attainable by citizens, social processes and organizations like PSM and their recommender systems should aim to meet truth-linked standards. Goldman (1987) distinguishes five epistemic standards by which to evaluate such systems and practices: (1) reliability, (2) power, (3) fecundity, (4) speed, and (5) efficiency. In the next section we elaborate these standards and how they relate to PSM recommender systems.

Epistemic welfare and PSM recommender systems

The epistemic standard of reliability refers to a system's ability to produce true beliefs and avoid errors (Thagard, 1997: 247). In the case of algorithmic recommender systems, this standard refers to their ability to ensure access to epistemically valuable content while avoiding epistemically poor or detrimental content. As part of the principle of distinctiveness, all content curated by PSM should fit the standard, while PSM recommender systems should promote content that is not only factually accurate but helps individuals make sense of their identities, histories, and societal roles. This includes shaping cultural knowledge, collective memory and social cohesion.

Power of a system constitutes its ability to deliver true answers to a diverse set of questions of user interest (Thagard, 1997: 247). Power here does not refer to social or political power but to how effective the knowledge practice is at answering many questions. In the context of PSM, the power of a recommender system is to effectively cater to a diverse set of user interests with a wide range of content representing various perspectives, befitting PSM's foundational principles of universality of appeal and distinctiveness. This increases the system's capacity to provide more knowledge generating answers, not just by distributing a wide range of content genres but also by including a diverse set of voices without racial, gender or ethnic bias (Fricker, 2015). This expands the knowledge base of the system, thus the system's ability to accurately answer more questions.

Fecundity relates to a knowledge practice's ability to maximize the number of individuals acquiring justified, true beliefs and is closely intertwined with the standard of power (Thagard, 1997: 247). In essence, fecundity refers to how easily a practice or system enables many people to answer questions of interest to them and to meaningfully make use of a system. This can be accomplished, for example, by designing interfaces that facilitate knowledge acquisition for a wide range of users. Systems that require specialized skills or background knowledge would not be considered fecund. For PSM's recommender systems, fecundity means guaranteeing universality of access by catering to citizens’ capabilities like digital literacy and familiarity with such systems. As such, it underscores the importance of accessibility and usability, ensuring that a PSM organization fulfills its role in serving the entire community within its remit. Fecundity relates to power when recommender systems provide cultural relevance and diversity, increasing its capacity to answer more questions and to introduce novel information while expanding audience engagement by reflecting more segments of society. This is about universality not only as equal access but as a commitment to epistemic inclusivity across diverse cultural, linguistic, and socio-economic contexts. In multicultural societies, this requires designing systems that accommodate differentiated epistemic needs while fostering shared civic knowledge, without assuming homogeneity in users’ prior knowledge, interests, or cultural affiliations.

Speed refers to the optimal rate at which content should be disseminated and processed to ensure accurate knowledge acquisition (Thagard, 1997: 247). Recommender systems for PSM must strike a balance between timely content delivery, delivery of relevant content, and the integrity of the knowledge it helps to cultivate. While rapid dissemination ensures that audiences remain engaged and informed, PSM's recommender systems should build in mechanisms that safeguard accuracy (timely error correction), depth, contextual richness and allow deliberative pacing.

Efficiency, finally, refers to the ability to limit economic costs of a system (Thagard, 1997: 247). In the case of a recommender system: to minimize the costs of distributing and accessing epistemically valuable content, reflective of removing barriers to universality of access. This extends to less visible financial burdens like environmental impact of AI systems and costs of personal data commodification. Commercial recommender systems are built to ensure profit maximization that can come at the expense of discovering or prioritizing epistemically valuable content or even economic efficiency. PSM, supported by public funding and guided by their foundational principles, should implement recommender systems that help citizens attain epistemically valuable states.

These epistemic standards can be thought of as evaluative criteria (Godler et al., 2019) to assess PSM content and services in general and their recommender systems in particular, that is, how well these systems perform on each standard. The more epistemic standards are met by PSM organizations and their recommender systems, the more conditions for citizens’ epistemic agency are created. This brings us to the governance by and of such recommender systems.

Epistemic standards and the governance by and of PSM recommender systems

Media governance is an essential component of fulfilling epistemic welfare and enhancing epistemic agency in the public sphere. On the one hand, recommender systems are a form of governance by algorithms (Just and Latzer, 2017). The epistemic welfare framework and standards can guide PSM's recommender system design. On the other hand, epistemic welfare can inform the governance of algorithms both in its prescriptive and evaluative capacity. PSM policymaking has notably been void of principles and guidelines for algorithmic systems like recommenders, in contrast to the governance of content, financial and other operational aspects of PSM organizations (Puppis and Ali, 2023). PSM organizations may be enabled and committed to develop noncommercial alternatives to platforms and their datafication-based business model (Cammaerts and Mansell, 2020). Given their foundational principles of universality and distinctiveness, it can be argued that PSM have a responsibility to develop and implement algorithms that differ from systems used by commercial players (Helberger, 2015; Tallerås et al., 2020; Van den Bulck and Moe, 2018). PSM responsibilities can also extend to the explainability and customizability of their recommender systems and to strengthening citizens’ epistemic agency by supporting development of relevant capabilities. Such obligations can be incorporated in PSM regulations, licenses and management contracts.

To show how this framework can help solve challenges related to PSM integrating recommender systems, we operationalize it into an actionable program, informing the design of PSM's recommender systems, providing evaluation metrics while inspiring policy-guidelines and tools for governance oversight. Each step, detailed in Table 1, illustrates how design, implementation and adjustment of algorithmic recommender systems require the continued participation of human actors. The framework's operationalization is informed by algorithm audit literature, adapted to the context of PSM to assess how well recommender systems uphold epistemic standards. Following Raji et al.’s (2020) approach, we first clarify the objective behind each standard: promoting the dissemination of and access to epistemically valuable content (reliability), that answers the greatest number of questions (power), to the broadest and most diverse audiences (fecundity), with minimal time (speed), and limited costs (efficiency). Building on these objectives, we follow Brown et al. (2021) in proposing assessment criteria framed as evaluative questions that, in turn, inform metrics for empirical evaluation as inspired by Zangerle and Bauer's (2022) overview of algorithmic evaluation aspects. Finally, as recommended by Raji et al. (2020: 41), we offer an “action plan” to help PSM organizations meet these objectives.

Table 1.

Framework: Application of epistemic welfare to recommender systems (RS) of PSM.

Epistemic standard	Reliability	Power	Fecundity	Speed	Efficiency
Objective	Promoting access to and dispersion of epistemically valuable content while minimizing circulation of epistemically poor or detrimental content (errors)	Fostering dissemination of and access to epistemically valuable content that answers the greatest number of questions	Maximizing dissemination of epistemically valuable content to the broadest and most diverse possible audiences	Minimizing time required to disseminate and access epistemically valuable content	Limiting resources to disseminate and citizens’ costs to access epistemically valuable content
Evaluation criteria	Does the RS provide accurate, fact-checked, and trustworthy content? Does the RS clearly label content (genre, e.g. news or opinion)? Is the RS dissemination unbiased? Are conflicts of interest in the RS's dissemination disclosed? Is the RS's accountability regarding epistemically poor or detrimental errors ensured?	Does the RS provide novel content? Does the RS provide societally/culturally relevant content? Does the RS prioritize epistemically valuable content? Does the RS provide content epistemically valuable to minority groups? Does the RS provide epistemically valuable content the user may not find otherwise?	Is the RS able to recommend to users with varying levels of education, digital/media literacy…? Is the RS interface easy to use and navigate? Is it easy for diverse users to find epistemically valuable content? Is the RS disseminating content that is diverse, has a variety of voices?	Does the RS deliver epistemically valuable content in a timely manner? Does the RS provide content with minimum steps? Does the RS quickly update, correct and adjust epistemically detrimental content?	Does the RS minimize (monetary, environmental, data/privacy) costs of dissemination and access? Does the RS disclose costs to users? Can users manage RS-related costs? Are costs justified relative to epistemically valuable content?
Evaluation metrics	Accuracy rate based on periodic audits. User surveys measuring perceived accuracy and trustworthiness. Number of corrections as measure of accountability. Transparency index rating (how is the RS’s functioning communicated to users).	Engagement metrics (e.g. click-through rates, dwell times) for content that answers user queries. User feedback on content relevance. Analysis of content for balance, diversity, and societally and culturally relevance.	User base diversity metrics on RS reach of various demographic groups. Usability scores from user experience surveys. Content analysis measuring representation of viewpoints, voices.	Latency measurements from query to content delivery. User ratings on timeliness of content received. Frequency and speed of epistemically valuable content updates in response to new developments.	User cost analysis (data usage, financial burden, time spent). RS environmental impact assessments. Third party services analysis. Accessibility measurements, incl. for economically and technologically impaired users.
Actions	Implement robust fact-checking protocols before dissemination. Clearly label content. Disclose conflicts of interest. Develop transparent error-correction policy and make it easily accessible to users.	Enhance content recommendation algorithms to prioritize relevance and novelty over commercial interests. Regularly update content libraries with current and comprehensive data. Use adaptive learning algorithms to tailor recommendations based on individual user interactions and societal relevance.	Design intuitive user interfaces, accessible to users with varying levels of digital literacy. Program the RS to actively recommend content reflecting minority perspectives and diverse voices. Ensure RS offers diverse content.	Optimize data processing and server response times. Implement streamlined RS navigation paths. Set up dynamic content updating system that does not compromise epistemic value of content.	Develop low-bandwidth version of RS for decreased data usage and cost for users. Communicate costs to users and provide options to manage these costs. Employ energy-efficient server technologies and optimize algorithms to reduce computational overhead.

Note. PSM: public service media.

The standard of reliability translates into evaluation criteria probing whether the recommender system disseminates epistemically valuable content in a balanced, transparent and trustworthy fashion. Evaluation metrics can include periodic audits for information accuracy, user surveys on perceived accuracy and trustworthiness, the number of corrections issued and transparency ratings, while potential actions to support these metrics involve robust fact-checking protocols, content labeling, disclosing conflicts of interest, and transparent error-correction policies (Diakopoulos, 2016; Shin, 2020).

The power of a recommender system enables users to find relevant answers to all their questions, offers novel content that prioritizes epistemic value over commercial or political interests, is societally and culturally relevant, includes minority perspectives, provides meaningful narratives and opportunities for engagement. Corresponding evaluation metrics can include implicit and explicit user ratings and content analysis for diversity (Vrijenhoek et al., 2021). Potential actions involve updating content regularly and adjusting (adaptive learning) algorithms to prioritize personal relevance and diversity (Heitz et al., 2022).

Fecundity considers the recommender system's ability to cater to both broad and diverse audiences, key to universality of access and appeal. Evaluation criteria involve the system's accessibility and ease of use, that enable a diverse array of citizens to navigate the system. Metrics might include user diversity, usability scores, and content balance analysis (Ekstrand et al., 2018). PSM can enhance fecundity through designing intuitive user interfaces (Nielsen, 1995) and promoting diverse perspectives (Fricker, 2015).

Regarding speed, the recommender system should deliver epistemically valuable content promptly, directly and accurately, including quickly correcting epistemic errors in recommendations and algorithmic designs. Recommender system can update too quickly, leaving users searching for previously seen content, or away from relevant content. Evaluation metrics include latency measurements, user ratings on timeliness and continuity, and frequency/speed of updates, while actions can involve optimizing data processing, streamlining navigation paths, and implementing dynamic content updating systems (Probst and Lommatzsch, 2016).

Efficiency of recommender systems means they minimize using economic resources and citizens’ costs. This requires accountability and transparency while ensuring sustainability, also in innovation. Metrics include user cost and accessibility measurements, third-party service analysis and environmental impact assessments, while actions may involve developing low-bandwidth versions of recommender systems, open-source coding methods, energy-efficient technologies and communicating costs to users (Sørensen et al., 2020; Zhou et al., 2024).

This framework approach aligns algorithmic recommender systems with epistemic standards, ensuring that dissemination, like content, follows PSM's foundational principles in a digitalized media environment. Its applicability is evident in recent PSM efforts to develop recommender systems. Carillon (2024) and Iordachi et al. (2025) show that actors at PSM organizations RTBF, VRT, and the BBC struggle to define “public service algorithms,” relying on disparate sources such as mission statements, policy mandates, peer practices, oversight, and scholarship. This underscores the pressing need for structured theoretical guidance. All three organizations lacked an operationalization and assessment of values, another gap the epistemic welfare framework addresses by proposing relevant metrics and actions.

These cases further illustrate how PSM organizations’ differences in remit, resources, governance and the media systems in which they operate affect how (well) recommender systems can be designed in line with epistemic welfare. RTBF and VRT's design considerations were guided by strong diversity mandates but faced resource constraints. At RTBF, this produced systems resembling commercial models. The BBC, while technically capable, prioritized engagement, with limited oversight allowing market pressures to dilute public service aims. Nonetheless, where mandates and organizational independence are robust, epistemic welfare interventions remain viable despite contextual influences which shape the framework's implementation yet not its normative strength.

Concluding remarks

To find ways for PSM organizations to design and implement algorithmic recommender systems that ensure their relevance while continuing commitment to their foundational principles and, thus, help overcome epistemic disruptions in the digitalized public sphere, we introduced the concept of epistemic welfare. Based on epistemic standards, we discussed ways to ensure dissemination of epistemically valuable PSM content that allow PSM to be part of the conditions and contribute to the capabilities that ensure citizens’ epistemic agency. As such, the framework can guide PSM policymaking through epistemic standards-based regulations and operationalizations in management contracts’ goals, KPIs, and metrics.

Some final considerations. First, our framework does not ignore or detract from the continued importance of human actors. On the one hand, human curation, editorial oversight and professional ethics are vital to shaping recommender systems, as part of creating and maintaining the conditions for epistemic welfare. As “the epistemic design of institutions” bears directly on acquiring justified true beliefs and knowledge (Godler et al., 2019: 216), algorithmic optimization always reflects the intentions and goals of the human actors that decide, guide and control their design and implementation. However, rather than assuming that human editorial judgment is inherently superior to algorithm decision-making, we argue that both benefit from prioritizing epistemic standards. As such, both human and algorithmic intervention should be guided by and evaluated against the epistemic standards of reliability, power, fecundity, speed, and efficiency for their contribution to users’ epistemic agency. This should be explicitly recognized as part of PSM's distinctiveness. On the other hand, our model is human centered in its attention to citizens’ agency. PSM responsibilities extend to the explainability and customization of their recommender systems to respect users’ autonomy and provide them with greater epistemic agency (Hildén, 2022; Piscopo et al., 2024). At the same time, it remains up to citizens to employ their agency and decide what they wish to prioritize. PSM could (be mandated to) help strengthen users’ capabilities, the complementary aim of epistemic welfare. Empowering citizens to strengthen their epistemic agency can be achieved by involving them in governance and by promoting their digital, media, data and information literacy, among other things. We agree with Cammaerts and Mansell (2020) that pushing the responsibility to ensure epistemic valuable states onto citizens can fall into the trap of the prevailing neoliberalist discourse. The epistemic welfare framework suggests that capabilities and conditions operate together: capabilities ensure citizens can take full advantage of the opportunities that the right conditions create. Advancing citizens’ epistemic agency in the digitalized public sphere can serve as a directional beacon for PSM to navigate between different responsibilities and develop digital strategies that seek to empower citizens.