Fact-checking in the multipolar AI order: Between epistemic sovereignty and ambivalence

---

Introduction: The crisis of fact-checking

Over recent years, fact-checking has experienced both a rapid rise in prominence and a growing crisis of legitimacy. Its expansion has been driven by the increasing significance of disinformation during major crises, including the COVID-19 pandemic and international conflicts. At the same time, doubts have intensified about whether systematic fact-checking can meaningfully contribute to political accountability or substantially influence news consumption across diverse audiences. On the one hand, the emergence of a professional community of fact checkers, represented by initiatives such as the International Fact-Checking Network (IFCN),¹See https://www.poynter.org/ifcn/ has helped address key challenges associated with the proliferation of disinformation. Shared methodologies, professional standards, and normative commitments have strengthened fact-checking as a corrective practice aimed at mitigating manipulative influence campaigns. On the other hand, fact-checking has increasingly become a broader discourse and practice of engagement with data, adopted by a wide range of actors who frame different forms of analysis as fact-checking. 

As Walter and Murphy (2018) highlight, empirical evidence on the effects of fact-checking is mixed. Meta-analyses indicate that although fact-checking can have measurable effects, these are generally modest and vary across different contexts (Walter et al., 2019). Its impact depends on audiences’ prior beliefs and political orientations. Fact-checks often fail to reach those most susceptible to misinformation, and even when they do, their corrective effects are limited by prior attitudes and partisan alignment (Nyhan & Reifler, 2010; Shin & Thorson, 2017). Individuals may reject corrections that threaten their social identity (Walter & Murphy, 2018), and even accepted corrections can leave “belief echoes” (Thorson, 2016). These dynamics emphasize that misinformation and its correction are not merely informational processes but also socially and politically embedded practices (Vraga & Bode, 2020). As Walter and Murphy (2018) highlight, empirical evidence on the effects of fact-checking is mixed. Meta-analyses indicate that although fact-checking can have measurable effects, these are generally modest and vary across different contexts (Walter et al., 2019). Its impact depends on audiences’ prior beliefs and political orientations. Fact-checks often fail to reach those most susceptible to misinformation, and even when they do, their corrective effects are limited by prior attitudes and partisan alignment (Nyhan & Reifler, 2010; Shin & Thorson, 2017). Individuals may reject corrections that threaten their social identity (Walter & Murphy, 2018), and even accepted corrections can leave “belief echoes” (Thorson, 2016). These dynamics emphasize that misinformation and its correction are not merely informational processes but also socially and politically embedded practices (Vraga & Bode, 2020).

Major international conflicts and crisis situations attract multiple actors who present competing assessments of information under the label of fact-checking. These claims often circulate widely and gain viral traction within ideologically aligned communities. In conflicts such as the Russia–Ukraine war and the Israel–Gaza war, each side effectively operates within its own fact-checking ecosystem. These parallel environments reinforce preferred interpretations while enabling audiences to dismiss alternative claims. For instance, in Russia, state-backed initiatives like the “War on Fakes”²See https://t.me/s/warfakes project and related Telegram channels use fact-checking language to selectively endorse claims that align with official narratives while dismissing opposing accounts as manipulation (Tuters & Noordenbos, 2024). Recently, fact checkers associated with the Iranian regime have highlighted protest images modified by AI as evidence that the protests themselves are staged by “Western propaganda” to delegitimize the protests themselves. (Alimardani, 2026). Consequently, individuals often become immersed in fact-checking bubbles that align with their political identities, a process further amplified by the partisan nature of some fact-checking initiatives that serve as a tool to delineate boundaries between political camps (Tsang et al., 2023). 

This trend is further intensified by authoritarian actors adopting fact-checking practices, as exemplified by the Russian state-sponsored Global Fact-Checking Network (GFCN)³See https://globalfactchecking.com/ (Reporters Without Borders, 2025). Alyukov (2024) argues that fact-checking has shifted “from a democratic initiative to a methodology for distorting reality” (online). Sharafutdinova (2020) suggests that, in Russia, identity politics act as a crucial tool for authoritarian legitimation by shaping an emotionally charged collective identity rooted in national pride, perceived external threats, and loyalty to leadership, particularly during times of uncertainty and insecurity. In authoritarian contexts, fact-checking supports identity politics by reinforcing pro-regime narratives (Shirikov, 2023) and protecting ontological security amid uncertainty and geopolitical volatility. Government-controlled fact-checking platforms also enable regimes to discredit criticism and justify repressive legislation, particularly in high-choice information environments where traditional censorship is insufficient (Alyukov & Zavadskaya, 2026). 

The function of fact-checking, therefore, should be considered in the context of how users engage with newsworthy content in high-choice social media environments. Although news consumption is often treated as an individual act, considering it through the lens of Bakhtin’s concept of multivoicedness and dialogism (Bakhtin, 1981) suggests that it can also be viewed as inherently dialogical. The dialogical nature of engagement with news is supported by the increasing role of social media as a site of news distribution, where information is consumed within the context of communication with other members of networks who share and comment on various newsworthy topics. In this light, news consumption can also be seen increasingly as an internal dialogical process in which individuals interpret information by engaging with a chorus of remembered and anticipated voices. From this perspective, fact-checking becomes one of the voices, or authoritative discourses, that participates in the negotiation of meaning. In the context of political conflict, fact-checking does not necessarily enter this inner dialogue as a neutral corrective. Instead, partisan fact-checking may strengthen the voice aligned with the reader’s preferred narrative and provide discursive resources for countering the imagined objections of opponents.

The authoritarian appropriation of fact-checking must therefore be situated within the broader rise of digital authoritarianism and the expansion of state capacities to deploy digital innovation for regime resilience. Against this backdrop, a key question emerges: how might new technological developments, particularly generative AI, either support fact-checking as a professional practice or enable new forms of authoritarian appropriation? Addressing this tension is essential for developing policy responses to an increasingly fragmented disinformation order (Bennett & Livingston, 2018).

Authoritarian appropriation of fact-checking resources: From the wisdom of crowds to artificial intelligence

Fact-checking relies on mobilising analytical resources to support critical engagement with socio-politically salient information. Traditionally, these resources have been provided by expert communities with specialised knowledge. Digital technologies have expanded this capacity by enabling crowdsourcing models that draw on collective intelligence (Allen et al., 2021; Martel et al., 2024). Crowdsourced fact-checking refers to participatory verification models that rely on distributed user contributions and leverage the “wisdom of crowds” to identify, evaluate, and contextualize potentially misleading information (La Barbera et al., 2024). While these models extend fact-checking capacity beyond experts, they typically operate alongside, rather than replace, professional fact-checking organizations. Tools such as Community Notes, a feature on X (formerly Twitter), enable participants to annotate potentially misleading content and add contextual information (Drolsbach et al., 2024).

At the same time, these participatory mechanisms are susceptible to political appropriation. Authoritarian actors have adopted this mode of fact-checking following the principles of “vertical crowdsourcing” (Asmolov, 2015) to promote regime-aligned interpretations. For instance, during the COVID-19 pandemic, the Russian state-sponsored project Coronafake engaged users to collect information on news reports that contradicted official accounts, including casualty figures and the effectiveness of the state’s response (Arkhipova & Peigin, 2021). Lapsha.Media, another Russian state-backed resource, provides users with a platform to share their analysis of what they believe are “fakes.” In this sense, crowdsourced fact-checking can be repurposed as participatory propaganda (Wanless & Berk, 2019; Asmolov, 2019), embedding verification practices within broader strategies of information control. The significance of state-backed participatory fact-checking initiatives lies less in the precise scale of user participation, which remains difficult to assess independently, than in their structural role within authoritarian information ecosystems: they harness participatory affordances for regime-aligned verification, operate as a form of state-sponsored digital vigilantism, engage users in the surveillance of dissent, and expand the scale of information control while suppressing or delegitimizing dissenting forms of participation (Gabdulhakov, 2021; Alyukov, 2022). This phenomenon is also evident within personal networks, where individuals assume fact-checking roles, creating tensions between interpersonal trust and ideological alignment. When beliefs central to political identity are challenged, users may sever ties rather than revise views. This reflects a form of disconnective power, defined as the state’s capacity to interfere in horizontal communication and reshape it in line with ideological priorities (Asmolov, 2018). Recent empirical research further indicates that political controversies, in which users advance competing interpretations of what happened by correcting, exposing, or denouncing others’ claims, are associated with significant spikes in public announcements of unfriending on Russian-speaking social media (Asmolov and Logunova, in press).

The rise of large language models (LLMs) marks a shift from collective intelligence and the “wisdom of crowds” to artificial intelligence and the “wisdom of machines” as a key analytical resource, including in fact-checking. Despite concerns about hallucinations and synthetic misinformation, LLMs show strong performance in verification-related subtasks (Augenstein et al., 2024). AI is therefore increasingly promoted as a scalable tool for verifying claims, detecting manipulation, and assisting fact-checking. Integration is already underway, mainly through hybrid human–AI systems rather than full automation. Organizations such as Full Fact use AI to detect claims and support verification workflows. Meta argues that AI operates as a second-opinion layer in fact-checking, enhancing the consistency of moderation without displacing human decision-making authority (Kaplan, 2025).

This shift marks not merely a technological upgrade but a transformation in the epistemic logic of fact-checking. Whereas crowdsourcing models rely on distributed judgment and visible contestation, AI-driven fact-checking concentrates epistemic power, the power to define what counts as credible or false, within opaque systems whose outputs are often treated as neutral or authoritative. This concentration creates new opportunities for political control, particularly in contexts where training data, model behaviour, and deployment environments are subject to state influence.

LLMs trained on censored or self-censored corpora may reproduce state-sponsored narratives (Ahmed et al., 2025). Bias embedded in training data can shape outputs in ways that reflect political constraints. For example, DeepSeek avoids sensitive Chinese political topics, often responding that such issues are beyond its scope (Mok, 2025). Bias is not confined to Chinese models. U.S.-based LLMs have produced distorted responses to politically sensitive queries posed in Chinese (Wiggers, 2025), and Russian LLMs such as GigaChat and YandexGPT heavily censor topics related to the Russia–Ukraine war and Chinese domestic politics (Meduza, 2025). Investigations by independent media outlets suggest reciprocal censorship patterns: Chinese LLMs restrict discussion of Russian politics, while Russian models censor China-related topics. This mutual alignment contributes to the joint production of sanitised realities across different national AI ecosystems. Journalistic experiments show that Russian LLMs often refuse to engage with sensitive topics by citing temporal restrictions or claims of neutrality, whereas Chinese models tend to reproduce official narratives more directly (Meduza, 2025). While existing evidence remains partial, these patterns illustrate how generative AI can reinforce authoritarian appropriation of fact-checking. Although these dynamics are most visible in Russia and China, they raise broader concerns as reliance on AI-driven fact-checking grows globally, including in Europe and the Global South.

Epistemological sovereignty meets AI sovereignty

Digital sovereignty has become a central concept shaping the governance of digital platforms across political systems. In authoritarian environments, it is framed as protection against foreign interference and regime destabilisation. In democratic contexts, it emphasises user autonomy and constraints on the power of major technology companies. Despite these differences, both approaches stress the role of the state in governing digital infrastructures to mitigate external threats (Pohle & Santaniello, 2024; Santaniello, 2026). Debates on AI governance represent a new phase in these discussions. The concept of AI sovereignty (Mügge, 2024), as the idea that states should retain control over the development, regulation, infrastructure, data, and cultural orientation of AI within their own borders, has gained traction among European policymakers and is increasingly invoked by China, Russia, and BRICS countries (an intergovernmental group of emerging economies originally comprising Brazil, Russia, India, China, and South Africa). Recent BRICS declarations adopted in Russia (2024) and Brazil (2025) call for sovereign national AI models to counter Western dominance (Asmolov, 2026). These initiatives frame sovereign AI not only as a technological project but also as a moral safeguard against epistemic inequality and cultural homogenisation. More recently, the Russian government began drafting legislation to mandate a ban on Western LLMs such as ChatGPT, Google Gemini, and Claude for users in Russia. 

Although the concerns of European policymakers are valid, governance models based on AI sovereignty risk fragmenting the AI landscape and creating national LLM bubbles. As Specht (2026) argues, the practical effects of AI sovereignty are already visible in the geopolitical fragmentation of AI infrastructure, as states increasingly treat data centers, compute capacity, and national AI capabilities as strategic assets to be funded, localized, and protected rather than as components of a shared global technological ecosystem. This trajectory reflects growing concerns about epistemic sovereignty, understood as the authority to define and legitimise knowledge within a political community (Oliveira & Pinto, 2024). When fact-checking becomes embedded in sovereign AI infrastructures, it ceases to be a peripheral media practice and becomes a core mechanism of epistemic governance.

In authoritarian contexts, sovereign AI projects embed state narratives directly into fact-checking systems, transforming fact-checking from a corrective practice into an instrument for delegitimising dissent. Government-aligned fact-checking organizations increasingly incorporate national AI models into their workflows, reinforcing official narratives. As Andrey Kondrashov, Director General of the Russian state news agency TASS, stated at the 6th Russia–China Media Forum in Beijing, “The GFCN uses its experience and AI technologies to identify false information and its primary sources.” Although this claim should not be read as independent evidence that such systems are fully operational, it is analytically significant as a performative statement of intent, showing how Russian state-aligned actors present AI-enabled fact-checking as part of a broader sovereign information infrastructure.

The impact of sovereign AI ecosystems, however, should not be reduced to the verified technical deployment of LLMs. It also reflects the discursive and institutional integration of verification systems within state-controlled epistemic environments, where influence operates through ecosystem alignment rather than tool adoption alone. Fact-checking organizations, media, and AI systems are thus increasingly imagined, and in some cases organizationally linked, within a unified framework that shapes how knowledge is produced and circulated, and stabilizes narratives across multiple layers of information production.

In a multipolar AI environment, even democratic states may pursue sovereign models that embed distinct epistemic assumptions. The result is epistemic fragmentation, in which competing AI-based fact-checking systems generate incompatible truths and undermine shared factual baselines. For citizens, this erosion of trust turns fact-checking into another arena of information warfare rather than a resource for democratic deliberation.

From fact-checking to AI-driven acceptance of ambivalence

The state of fact-checking today mirrors a wider epistemological crisis, worsened by geopolitical tensions, political polarization, and the increasing influence of AI-driven applications. Its vulnerability to authoritarian appropriation highlights the need to reconsider fact-checking as a form of communicative action involved in constructing socio-political realities (Andersen & Søe, 2020). This does not mean abandoning truth claims but rather questioning whether binary truth arbitration can effectively address disinformation in highly polarized and sovereign information environments. 

Bauman’s concept of ambivalence provides a productive lens for this reconsideration. Ambivalence refers to the coexistence of multiple, conflicting interpretations that resist clear classification (Bauman, 1991). Modern institutions attempt to eliminate ambivalence through rigid distinctions between true and false, yet such efforts inevitably fail as social reality exceeds imposed categories. From this perspective, fact-checking represents a modern attempt to manage uncertainty by enforcing epistemic closure on contested information.

Importantly, ambivalence should not be conflated with relativism. Recognising the coexistence of competing interpretations does not imply that all claims are equally valid or that evidence is irrelevant. Rather, it acknowledges that in complex political conflicts, factual claims are embedded in broader narratives, moral frameworks, and identity commitments that cannot be resolved through verification alone. Orgad (2025) argues that contemporary media culture suffers more from excessive certainty than from doubt. She calls for greater acceptance of incompleteness, discontinuity, and lack of closure in public discourse (Orgad, 2012). New technologies, she suggests, create opportunities for more ambivalent media spaces in which events may be assigned to multiple categories or resist categorisation altogether (Toussaint, 2014). Rather than treating ambivalence as a failure, such spaces recognise it as an unavoidable condition of political life.

Generative AI is particularly well-suited to supporting this form of engagement. LLMs can present the same event from multiple viewpoints, enabling users to explore not only what happened but how different social, cultural, and political positions shape interpretation. In this sense, AI can support a shift in the role of fact-checking, moving it away from reinforcing the reader’s preferred narrative and towards facilitating dialogic imagination, understood as openness to multiple voices rather than reliance on a single authoritative interpretation (Bakhtin, 1981). Reframing AI’s role from an automated fact-checker to a facilitator of perspective-taking reduces its susceptibility to authoritarian appropriation by shifting the emphasis away from epistemic closure and toward epistemic reflexivity.

Rather than positioning AI systems as arbiters of truth, platforms and fact-checking organisations could deploy them as tools that surface competing interpretations, contextualise claims, and make visible the assumptions underlying different narratives. Such an approach would align fact-checking more closely with democratic deliberation by supporting critical reflection rather than enforcing agreement. This reorientation also opens space for identifying common denominators across antagonistic narratives. Orgad (2025) suggests that human suffering may function as a meta-frame present across opposing accounts of the same event. Rather than resolving disagreement through classification, AI-supported engagement with ambivalence can help audiences recognise shared concerns while acknowledging persistent differences.

This approach can be seen as emerging from the convergence of multiple existing project types, each contributing elements that could support its development. The first type of project relies on user engagement to ensure that diverse points of view are represented while adding contextual information to contested posts. This is represented by Community Notes on X. According to the project’s statement, “Community Notes takes into account not only how many contributors rated a note as helpful or unhelpful, but also whether people who rated it seem to come from different perspectives.”⁴See https://communitynotes.x.com/guide/en/contributing/diversity-of-perspectives The second type of project, such as Polis,⁵See https://pol.is/home demonstrates how computational systems can map areas of agreement and disagreement, identify consensus and divisive statements, and make the structure of collective opinion visible. These initiatives are particularly important, as the classification structure is not necessarily imposed but may emerge in a bottom-up manner from the mapping itself. The third type of project includes news-contextualisation platforms, such as Ground News⁶See https://ground.news/ and AllSides,⁷See https://www.allsides.com/unbiased-balanced-news which organise news narratives around multiple perspectives and ideological positions to make media bias visible. However, these projects rely on a rigid categorisation of perspectives that mainly distinguishes among left, centre, and right positions in the media. In contrast, emphasizing ambivalence and multivoicedness indicates that relying on strict structures to classify multiple perspectives should be approached carefully, as it might inadvertently reinforce divisions. These three groups of initiatives should not, however, be understood as examples directly linked to deliberative fact-checking aimed at mitigating the risk of AI-based epistemological fragmentation. Rather, they indicate a broader shift towards infrastructures that can contextualise claims, make disagreement visible, and support human interpretation. The article’s proposal remains more normative and more ambitious than most existing systems, as it focuses on how AI could help address the fragmentation of fact-checking by harnessing ambivalence as an organising principle of engagement with information and making multivoicedness more visible.

A key challenge in operationalizing this approach is distinguishing legitimate interpretive plurality from the amplification of falsehoods or bad-faith narratives. Ambivalent media should not abandon categorization but pluralize it, shifting from binary classifications to layered epistemic categories that distinguish between verifiable claims, contested interpretations, and underlying assumptions. This reorientation can be implemented across multiple layers of AI systems. At the model level, it involves training or fine-tuning systems to represent structured disagreement rather than converge on single authoritative outputs. At the interface level, it is realized through design features that surface competing perspectives, contextualize claims, and signal uncertainty.

This approach must also reflect how audiences process information, as reliance on heuristics may limit engagement with complexity. Rather than presenting singular narratives or binary outcomes, AI-supported ambivalent interfaces can operationalize ambivalence through structured designs that make uncertainty and plurality cognitively manageable. For example, they could offer layered responses that separate verified facts from contested interpretations and highlight uncertainty, making complexity navigable while preserving distinctions between disagreement and misinformation. Such epistemic scaffolding can align with familiar modes of information consumption. Intelligent user interfaces (IUIs) provide a pathway for operationalizing this approach by simulating debates between recognizable viewpoints. This externalizes the interpretive process and makes multivoicedness guided rather than open-ended, enabling users to engage with contested realities without cognitive overload. Autonomous AI agents may further structure interactions between competing perspectives, although this remains an open area for research.

Such an approach does not deny the importance of factual accuracy. Instead, it recognises that classification alone cannot resolve the complexity of social reality in conditions of conflict and polarisation. By helping users navigate ambivalence with greater awareness, empathy, and reflexivity, AI can support more resilient forms of public engagement with contested information. In doing so, it offers a pathway for preserving the democratic value of fact-checking while limiting its authoritarian appropriation in an increasingly fragmented epistemic landscape.

The proposed shift in AI-supported fact-checking should be viewed as exploratory rather than prescriptive. While generative AI may support multivoiced engagement, this potential remains largely untested and depends on design choices and user interaction. Therefore, the role of AI in engaging with contested realities should be treated as a research agenda requiring further empirical study, including experiments, interface testing, and user-centred evaluation. Such work is needed to assess whether ambivalence-oriented approaches can enhance critical engagement without reinforcing relativism, while also mitigating the risks of authoritarian appropriation.