The algorithmic public opinion as an open heuristic

There is little doubt that the persisting integration of algorithmic systems into public opinion formation processes continues to spark lively and cross-disciplinary debate. Issues are still raised around the interplay between human and technological actors and the evolving power dynamics that shape the visibility, circulation and interpretation of digital content. In this response, we welcome the opportunity to engage with the insightful commentaries provided by Bodo (2025), Bruns (2025), Lim (2025), Kant (2025), and Prodnik (2025) all of whom have raised important points not only on our choice of terminology and conceptual framing, but also on broader scholarly concerns such as user agency, technological determinism, and the broader political economic dynamics that we invoke when constructing the notion of ‘algorithmic public opinion’. We are grateful to them and their careful engagement with our work. Their reflections present us with a valuable opportunity to clarify, nuance, and extend our thinking, as well as help us further situate the concept we propose within such a rapidly evolving setting.

The commentaries highlight both convergences and divergences in how scholars interpret algorithmic mediation of public opinion. They remind us that concepts like the one we propose are never neutral descriptors, but interpretive tools shaped by disciplinary perspectives. In this response, we aim to address six clusters of comments that emerged across the contributions: the heuristic function of the notion of ‘algorithmic public opinion’; the socio-technical and non-deterministic role of algorithms; the qualitative shift in public opinion formation; the multifaceted nature of user agency; the interplay of social media, financial resources, and political power; the implications of AI and emerging technologies. We hope this exchange contributes to advancing the study of the conceptual, empirical, and normative challenges involved in algorithmically mediated public opinion processes – which, we contend, is crucial especially as we move into the uncharted territory of generative AI tools and their mass availability.

Argument #1: ‘The algorithmic public opinion as a reductive label’

In our article, we argue that the integration of algorithmic systems into large online social networks since the early 2010s has given rise to an ‘algorithmic public opinion’, a novel form of public opinion formation whose distinctive feature lies in the centrality of automated decision-making processes and their interaction with a variety of actors. These interactions shape which issues of public interest gain salience, what news sources and interpretations prevail, and how patterns of news consumption evolve. This concept is designed to highlight how social media algorithms intervene in the processes of making and shaping public opinion, offering a framework for studying both the processes that shape opinion and the dynamics that emerge from opinion makers and publics.

In his commentary, Bodo (2025) rightly observes that the field of critical algorithmic studies has often relied on broad, sometimes reductive terms to capture what are instead highly complex socio-technical dynamics. He raises concerns about the term ‘algorithmic public opinion’ (APO, from now on) reproducing such a shorthand. It is important to acknowledge this potential, and we share the view that academic terminology must be treated with caution. Labels, especially somewhat ‘catchy’ ones, can both illuminate and obscure the objects of study. We agree with Nick Seaver (2017) when he writes: ‘I take terminological anxiety to be one of critical algorithm studies’ defining features’, adding that these are ‘first and foremost anxieties about the boundaries of disciplinary jurisdiction, and critical algorithm studies is, essentially, founded in a disciplinary transgression’ (p. 2).

When proposing this notion we did not want to offer another definitive or reductive label, but rather delineate the boundaries of an inquiry that has extended in a variety of directions. Besides algorithms, the notion of public opinion is also broad, and varies depending on the disciplinary area, as we discuss in the paper. We determined that having a single concept would help researchers situate the scholarly fields we were going to move on. As we write the notion of APO is conceived as a pragmatic device to capture, in a single term, the manifold ways in which social media algorithms intervene in public opinion formation in contemporary digital societies (Gandini et al., 2025: 11). While heuristics are simplifying tools, our aim is not to over-simplify, but rather provide an intuitive reference that facilitates comparative work and offers a potentially shared framework to approach similar questions from different methodological and theoretical standpoints.

Similarly, Bruns (2025) reminds us that ‘terminology matters, as it directs us towards a specific conceptualisation of the phenomena it describes, and encourages a particular operationalisation of these concepts in subsequent empirical research’ (p. 4). We fully agree, and we see APO not as a final conceptual endpoint, but as an invitation to further theoretical development and empirical operationalisation. No single term can capture the full complexity of the objects of study, but APO can provide a starting point – that is inevitably provocative and open to critique and extension – to address and continue addressing the evolving role of social media platforms in shaping public discourse.

Argument #2: The algorithmic public opinion is ‘algorithmic-centric’ and ‘deterministic’

Bodo (2025) cautions that the concept of APO risks failure because ‘it hinges its fate upon the term algorithm’, noting that algorithms are not autonomous agents but socio-technical assemblages involving software, data, designers, operators, and owners (p. 2). We agree with this premise. If parts of our text suggested otherwise, the responsibility lies with us for not being sufficiently clear. Our intention was precisely to emphasize that algorithmic systems are embedded in broader socio-technical ecologies, shaped by both platform policies and the actions of multiple actors. This is why we present APO as a framework for understanding the central role of automated decision-making processes in shaping the circulation of information in interaction with human actors (Gandini et al., 2025: 2). Our distinction between direct algorithmic gatekeeping (via moderation and personalization) and indirect algorithmic gatekeeping (through the actions of users, media, and institutions) is meant precisely to underline this hybridity, rather than to suggest determinism.

Bruns (2025) similarly warns against ascribing ‘irresistible power’ to algorithms, noting that they may channel opinion formation but do not determine it exclusively (p. 3). We share this concern as we write that algorithms ‘contribute’ to shaping public opinion (Gandini et al., 2025: 2), thus carefully framing their role as partial rather than absolute. We foreground the entanglement of automated decision-making with human agency in two key areas of our paper. First, we stress the importance of understanding algorithmic public opinion as a product of interactions between users and algorithmically curated informational content, stressing how users actively negotiate their agency within these ecosystems, rather than being passive recipients of algorithmic influence. Second, through the term ‘indirect gatekeeping’ we shed light on the role of intervening actors who seek to shape public opinion by leveraging, interpreting, or even speculating on the mechanisms of algorithmic systems. This underscores that human intentionality is deeply embedded within our understanding of algorithmic formation of public opinion.

At the same time, as other commentators (Kant, 2025; Prodnik, 2025) note, we share the view that algorithms have the capacity to drive, via their infrastructural nature, the ways in which public debates develop. Content moderation practices around military conflicts aptly illustrate this dynamic: while moderation policies are ultimately human decisions, they are enforced at scale through algorithmic infrastructures that are opaque, partly unpredictable, and often contested. The suppression of pro-Palestinian content on Meta platforms, documented by Human Rights Watch (Brown, 2024) and Access Now (Fatafta, 2024), shows how algorithmic infrastructures can decisively shape the visibility of political expression. Users’ resorting to ‘algospeak’ is itself a recognition of the structural role algorithms play in mediating what circulates publicly (see Aleksic, 2025). It is in this sense that algorithms, while not autonomous agents, nonetheless acquire a significant role in structuring the conditions of public opinion formation – though partially, as said, rather than ‘absolutely’.

Argument #3: ‘Algorithms and the qualitative change in public opinion formation’

Bruns (2025) invites us to be cautious about treating APO as evidence of a ‘fundamental state change’ (p. 11). He raises the important question of whether algorithms constitute a genuinely epochal shift, or whether they simply reconfigure existing dynamics by reducing the influence of legacy media while amplifying that of digital platforms. We agree that this distinction is crucial. APO is not meant to suggest that algorithms alone represent a rupture. We see them as central features of a broader process of transformation of communication, information flows, and public life at large (Floridi, 2014). In this sense, algorithms are not isolated drivers of change, but key infrastructures within a larger socio-technical transformation that is still unfolding. Isolating their distinctiveness, we contend, is crucial to analysing the present conditions of public opinion formation – which may further evolve as artificial intelligence becomes increasingly embedded in these infrastructures (more on this later). Although it is always difficult to anticipate the future, it seems reasonable to expect that the production, circulation, and reception of news and opinions will continue to shift in ways that make the contemporary environment qualitatively different from the pre-digital age. The notion of APO, in our view, is a helpful way of naming and probing this environment.

To capture dynamics more cautiously, Bruns (2025) suggests that a term such as ‘algorithmically-informed public opinion formation’. We understand the motivation behind this proposition. However, we ultimately opted in using the term algorithmic public opinion, as it grasps the interaction between algorithms, platforms, and publics – the ‘complex interplay’, in our words. Following this reasoning, APO recognises a central but not deterministic role to algorithms, while recognising the entangled dynamics and co-creative role of all actors participating. Instead, ‘algorithmically-informed’ public opinion suggests a unidirectional relationship in which algorithmic outputs merely shape users’ opinions.

Argument #4: ‘User are passive and manipulated, or free and active agents?’

Several commentators have raised questions about how we describe user agency in our article. Bruns (2025) contends that users ‘retain significantly more agency than this article's discussion affords them’ (p. 2), while Lim (2025) emphasizes the role of everyday practices, arguing that ‘vernacular visibility modulation’ positions users not as passive recipients but as cultural agents whose actions, however mundane, help shape publics (pp. 4‒5). Conversely, Bodo (2025) and Prodnik (2025) caution against overstating agency, highlighting the structural pressures of algorithmic systems that can habituate audiences and create compulsive patterns of behaviour (Bodo, 2025: 4; Prodnik, 2025).

We interpret these divergent readings as indicative of both the complexity of user engagement with algorithmic systems and the challenges of articulating its extent clearly. Our article intentionally refrains from reiterating dichotomies (such as ‘users are passive’ vs. ‘users have agency’) and instead stresses that audiences are hyper-heterogeneous. As we write, they are ‘inherently diverse, encompassing users from different generations, varying levels of educational attainment’, which translates into ‘highly heterogeneous behavioural patterns extrapolated through algorithmic means, as well as varying modes of interaction with and influence over algorithms’ (Gandini et al., 2025: 11). Importantly, we underline that user agency can also manifest through forms of resistance, as highlighted by Bonini and Trere (2024) among others; users may subtly subvert, circumvent, or creatively repurpose algorithmic infrastructures, demonstrating that agency is not only constrained or passive but also active and strategic. Lim's (2025) observation that ‘affect is not a mere byproduct of algorithmic circulation but the connective tissue that binds users, content, and systems within continuous feedback loops’ (p. 4) further reinforces this view – that agency emerges not only through behaviour but also through affective and relational entanglements with content and systems. We see this as a fruitful avenue for further conceptual and empirical exploration, complementing APO's focus on the infrastructural and algorithmic dimensions of public opinion formation.

Argument #5: ‘The interplay of media, power, and politics’

Bodo (2025) observes that ‘the relationship of media and public opinion is in some sense the history of the relationship between political power and the power of media products and owners’ (p. 2), and that ‘the focus on the commercially driven algorithm masks the fact that these algorithms are, first and foremost, tools in the hands of the media companies to reach their own long-term political goals, which are much broader than just making more money’ (p. 3). Prodnik (2025: 4) similarly notes that platforms ‘act as vast lobbying entities’ whose interests increasingly overlap with those of dominant states. We fully agree with these observations that highlight important dimensions of power that our article did not emphasize sufficiently. Yet, this is not necessarily new either. In particular, actors with greater financial resources – including legacy media companies – have always had a disproportionate capacity to shape public opinion. This is not a novel feature of algorithmic systems but reflects longstanding inequalities in media influence. At the same time, the notion of APO was not intended to serve as a ‘genuinely critical framework’ in the normative or ideological sense in relation to the political economy of digital media. Rather, it was conceived as an analytical tool to understand how algorithmic dynamics concur to shape public opinion in the digital era, and the complex web of actors that intervene within this setting. Its focus on the interplay between algorithms, digital media infrastructures, and publics, we contend, provides a conceptual lens to understand their entanglement without necessarily claiming to offer a political or normative critique – albeit arguably much needed and welcome.

Argument #6: ‘Algorithmic public opinion in the age of AI’

Kant (2025) rightly points out that generative AI ‘should be considered as capable of intervening in and indeed constituting algorithmic public opinion’, noting that systems such as LLMs mediate interactions primarily between individual users and AI-generated content rather than between users themselves (pp. 3‒4). We fully agree. This article has been quite long in the making, and its genesis precedes the mainstreaming of generative AI. Thus, having come at the final stage of review, we made the conscious decision to avoid opening a new line of inquiry in an already dense conceptual piece. However, it is important to clarify that the ‘algorithmic’ in APO encompasses a broad spectrum of AI-driven systems. Recommender systems themselves are not static algorithms but machine learning infrastructures. While generative AI and other AI systems are indeed likely to further reshape the production, dissemination, and consumption of information in ways that are still largely unpredictable, they can also potentially be understood as additional layers in the automated circulation of content within our framework, thus contributing to APO mainly as a form of indirect gatekeeping. These also share several of the key features that characterize the systems we have already discussed in our article: (semi)automation, speed and scale of operation, opacity and inscrutability, and the existence of socio-technical feedback loops that continuously reshape both content and user behaviour.

AI's influence already spans multiple dimensions: from conversational search engines to low-quality AI-generated content (so-called ‘AI slop’) (Stanusch et al., 2025). These developments underscore that APO is evolving alongside such technological shifts, rather than becoming obsolete. Even as generative AI systems add a distinct layer by operating largely through user-driven interactions, they are embedded within the broader infrastructure of automated content curation, ranking, and moderation. For example, AI slop, rather than being an accidental by-product of generative technologies, reflects similar engagement-driven dynamics that govern social media's recommender systems. Low-quality ‘brainrot’ content produced at scale is precisely what tends to be algorithmically amplified. In this sense, all these systems are mutually entangled, continuously influencing each other and contributing to the formation of public opinion altogether. Prodnik (2025) highlights the reciprocal relationship between platforms’ investments in generative AI and the extraction of public opinion as raw material. This also points to the intertwining of APO and AI-driven infrastructures: the same mechanisms that shape public opinion also feed into AI systems, reinforcing the importance of a unified conceptual framework.

Looking forward, we may forecast the rise of ‘conversational recommender systems’ that may offer new forms of user interaction, allowing audiences to express preferences directly and influence content recommendations in real time (Lazar et al., 2025). Should such systems replace current social media's recommender systems, they might fundamentally alter or even disrupt the dynamics that define APO as we describe it, giving rise to a qualitatively different APO in which users directly shape algorithmic logics. Yet, we conceptualized APO precisely because it has become institutionalized, extending its influence into cultural production itself (Poell et al., 2021). Societies and users are now dependent on the logics of APO not only economically or organizationally, but even psychologically. Most people are used to, and even find comfort in, the rhythms and rewards of algorithmic social media engagement. Any substantial transformation that might occur will take place while the infrastructures and habits that sustain APO are still deeply embedded in contemporary social life. While the notion of APO may need to evolve to incorporate these changes, we argue that its conceptual kernel – highlighting the role of automated and algorithmic systems in structuring public opinion – remains valuable.

Conclusion

The commentaries on our article have highlighted important questions and critiques regarding terminology, algorithmic determinism, user agency, the interplay of media and power, and the role of artificial intelligence. Engaging with these perspectives has been refreshing and has reinforced our view that the notion of APO may be a valuable heuristic if seen as a flexible tool – rather than a finalized concept – to grasp the complex and evolving ways in which algorithmic and AI-driven infrastructures contribute to shaping public opinion formation. We hope that the dialogue initiated by these commentaries can extend the debate on how algorithmic and AI-driven systems interact with publics, media, and political and economic forces further, and that APO can effectively serve as a productive framework for better understanding the continuously shifting dynamics of public opinion in the digital age, while remaining responsive to new technologies, societal changes, and critical perspectives.