Fact-checking arrived as a bright promise: a way to cut through the noise of a chaotic information age, to give citizens reliable truth when lies were spreading faster than ever. During the pandemic and major conflicts, fact-checking grew rapidly, and professional networks with shared standards emerged to meet the threat of disinformation. But alongside that growth, a deeper crisis has unfolded. Study after study shows that fact-checking has only a modest effect on what people believe. Those most likely to fall for misinformation rarely see the correction, and even when they do, they often reject it because it conflicts with their existing identity and political loyalties. Worse, even when people accept a correction, it doesn’t fully erase the original false memory—it leaves behind what researchers call “belief echoes.” The deeper problem is that misinformation and correction are not just about facts and data. They are deeply social. People consume news not as isolated individuals but as participants in a crowded mental conversation, influenced by the voices of friends, pundits, and political leaders. Fact-checking does not enter this conversation as a cool, neutral referee. It enters as one more voice, sometimes welcomed, sometimes mocked, but always filtered through the listener’s existing worldview.
This is why fact-checking has become so easy to weaponize in polarized times. Major conflicts now come with competing fact-checking ecosystems on each side, each producing its own certified truths and debunked falsehoods. In the Russia–Ukraine war and the Israel–Gaza war, for example, each side’s fact-checkers work to reinforce preferred narratives and to dismiss the other side’s claims as manipulation. Authoritarian states have been especially quick to learn this game. Russia’s “War on Fakes” project uses the language of fact-checking to endorse official narratives and to label independent reporting as enemy propaganda. Iranian state-aligned fact-checkers have used examples of AI-manipulated protest images to claim that the protests themselves were staged as Western propaganda. The result is a world of fact-checking bubbles. People live inside ecosystems that validate their own side and make the other side’s arguments easy to dismiss. This is not an accident. It is a deliberate strategy. For authoritarian regimes, fact-checking has become a tool of identity politics, building national pride and loyalty by defining an “us” against a threatening “them.” It protects people’s sense of security in an unstable world, but it also draws harsh lines between political camps. Fact-checking is no longer just about correcting falsehoods; it is about delineating who belongs and who is the enemy.
For a long time, fact-checking relied on two kinds of resources: the expertise of journalists and scholars, and the collective intelligence of the crowd. Crowdsourcing models, like Community Notes on X, allowed ordinary users to add context and flag misleading content. This seemed like a democratizing step. But authoritarians were quick to adapt. They created their own version of crowdsourced verification, sometimes called “vertical crowdsourcing,” where the state invites citizens to report “fakes”—meaning any news that contradicts the official line. During the pandemic, Russian state-backed projects such as Coronafake enlisted users to submit reports of inconvenient news, from casualty figures to questions about government response. Platforms like Lapsha.Media encouraged citizens to share what they considered disinformation. This turned fact-checking into a form of participatory propaganda, a way to mobilize citizens as digital watchdogs for the regime. It wasn’t exactly participation in the democratic sense; rather, it was a state-sponsored form of digital vigilantism, drawing people into the surveillance of dissent and expanding the reach of information control without ever appearing heavy-handed. Now, with the rise of artificial intelligence, the game is changing again. The “wisdom of crowds” is giving way to the “wisdom of machines.” AI systems, especially large language models, can process huge amounts of information at scale, identify false claims, and support verification. But these machines are not neutral. They are trained on data that can be censored, biased, or shaped by political pressure. Chinese and Russian language models, for example, have been shown to avoid sensitive topics or to reproduce state official narratives. Russian models refuse to discuss certain aspects of the war in Ukraine; Chinese models are careful about both Chinese domestic politics and Russian politics. In a joint study, independent media found a pattern of reciprocal censorship. This is not just a technical flaw. It means AI-based fact-checking, once it is integrated into national information systems, can become a powerful instrument for authoritarian regimes to define and enforce their own version of reality.
The push for “AI sovereignty” is making this problem even more acute. Around the world, governments are demanding their own national AI models—systems that are not dependent on American or Chinese tech giants. In many cases, this is framed as a matter of security, culture, and self-determination. But it also carries a deeper implication: epistemic sovereignty, or the right of a nation to control what counts as knowledge within its borders. Russia has talked about banning Western chatbots like ChatGPT; China develops its own models with strict content controls; and BRICS declarations call for sovereign AI systems to counter Western dominance. On the surface, these initiatives claim to protect against epistemic inequality and cultural homogenization. But the practical effect is the fragmentation of truth itself. When each country has its own AI fact-checking system trained on its own state-approved corpus, these systems will generate incompatible truths. A claim that is verified in one country will be dismissed as disinformation in another. Shared factual baselines, already fragile, collapse entirely. Authoritarian states are particularly eager to embed their narratives directly into AI fact-checking workflows. When state-aligned organizations incorporate national AI models into their verification processes, fact-checking no longer checks facts; it becomes an instrument of epistemic governance, a way to legitimize state-sanctioned narratives and to systematically delegitimize dissent. Even democracies are not immune. The more they invest in national AI models, the more they risk isolating themselves inside their own knowledge bubbles. The result is a multipolar world of competing, mutually exclusive realities, where fact-checking becomes just another arm of information warfare rather than a resource for public reason.
Given all this, it is tempting to ask whether fact-checking can be saved at all. Perhaps the obsession with separating truth from falsehood in a simple binary is the heart of the problem. The sociologist Zygmunt Bauman used the word “ambivalence” to describe the condition where multiple conflicting interpretations coexist and resist easy classification. Modern institutions hate ambivalence. They want clean rules, clear verdicts, a true or false button. But social reality is too messy for that. In complex conflicts, factual claims are tangled up with narratives, moral frameworks, and identity commitments. You cannot resolve a war through fact-checking any more than you can resolve a marriage by arguing over who said what first. This doesn’t mean everything is relative—it doesn’t mean lies are just opinions. It means we need a different way of engaging with contested information, one that acknowledges multiple viewpoints and helps people navigate them. This is where AI could actually help. Instead of using language models as final arbiters of truth, we could use them as guides to multiple perspectives. Imagine an AI system that, when asked about a disputed event, shows you not only the verifiable facts but also how different communities interpret them. It could surface the underlying assumptions of different narratives, making visible why people on opposite sides believe what they believe. This is what the Russian philosopher Mikhail Bakhtin called multivoicedness, and what others call dialogic imagination. A fact-checking system like this would not try to close down conversation with a single verdict. Instead, it would open up a space for reflection, helping people see the world through more than one set of eyes. There are already small glimmers of this approach. Community Notes on X tries to include notes from people of different perspectives. Polis maps consensus and disagreement in public opinion. Ground News and AllSides show the same news story from left, center, and right media. These projects are not perfect; they often rely on rigid categories that can deepen divisions. But they point toward a different possibility: fact-checking as a tool for understanding, not just for ruling.
Of course, this idea comes with serious challenges. We cannot simply treat every claim as equally valid or give a platform to bad-faith lies. The point is not to abandon truth, but to pluralize it, to move from a binary of true and false to layers of verification: what is proven, what is contested, what is uncertain. We also need interfaces that make this complexity manageable. Having ten conflicting interpretations on the screen doesn’t help anyone. AI systems could be designed to present layered responses—verified facts first, then competing interpretations, then the assumptions behind those interpretations. They could simulate debates between recognizable viewpoints, pushing users to engage with opposing arguments without overwhelming them. But this is not a recipe that can be applied overnight. It is a research agenda, one that needs careful experimentation, user testing, and a genuine commitment to democratic values. We need to ask tough questions: Can an ambivalence-based system avoid reinforcing the very polarities it tries to reveal? Can it help people tolerate uncertainty rather than seek comfort in certainty? The great danger today is not too much doubt, but too much certainty. We are drowning in people who are absolutely sure they are right, and that certainty is often manufactured by political forces using the tools of fact-checking and AI. If we can build fact-checking systems that teach people to live with ambiguity, to listen to other perspectives, and to hold their own beliefs with a little more humility, then we might preserve the democratic promise of fact-checking after all. Not as a machine that hands down verdicts, but as a mirror that shows us ourselves, and sometimes, the face of the other. That is a harder and more hopeful task. It asks us to be not just fact-checkers, but citizens.