Why What We Know About Polarization Is Mostly American

TURN off Facebook for a month in the United States, and strangers on opposite sides of the political aisle warm to each other slightly. Run the identical experiment in Bosnia-Herzegovina, and polarization gets worse, not better. That reversal is not a fluke: a new review in Nature Communications argues that most of what social science thinks it knows about political polarization comes from a single country, and evidence gathered elsewhere often runs in the opposite direction. Its authors, led by King’s College London‘s Max Falkenberg, went looking for exactly how lopsided the evidence base really is.

Falkenberg and his co-authors combed the literature and counted: 5,965 papers use the terms “political polarization,” “ideological polarization,” or “affective polarization” in their title or abstract, and 2,243 of those name a specific country. Of the papers that do name a place, 967 of them, 43.1%, are about the United States, more than any other nation by a wide margin.

The imbalance matters because polarization does not mean the same thing everywhere. Merriam-Webster made “polarization” its word of the year in 2024, defining it as a society splitting into two sharply distinct camps, a framing that fits a two-party system neatly enough. European democracies routinely run on coalitions among half a dozen or more parties, and researchers studying them are still arguing over whether the same two-camp logic even applies. New Zealand offers an unusually clean natural experiment in how much electoral rules alone can shape the answer: when the country switched from a winner-take-all system to proportional representation in 1993, the measured ideological polarization of its political elites doubled within a few years.

A voting-system change is one thing; the internet and artificial intelligence are another, and here the review’s central tension sharpens. Because the platforms hosting most political conversation, and the biggest randomized experiments testing what they do to us, are almost all American.

The Experiments Almost Nobody Runs Abroad

The clearest causal evidence anywhere in this field comes from a series of experiments run in 2020, timed to that year’s US election, with Meta’s cooperation. The largest of them found that cutting people’s exposure to politically like-minded sources on Facebook did not measurably change their affective polarization or ideological extremity. That collaboration, though, was granted only to US-based researchers studying US users, what one commentary termed independence by permission. Nobody has run an equivalent study on, say, X in Brazil or WhatsApp in India, the platforms actually doing the political work across much of the world.

“Social science has an America problem,” says Falkenberg, an AI+ senior fellow at King’s and the review’s lead author, pointing to norms established in the US, including the assumption that political antagonism sits on a single left-right axis, that “don’t necessarily fit with what’s happening in other parts of the world.” The election of Donald Trump in 2016, and the social media campaign that carried it, he says, is what “caused an explosion in polarisation studies in the US that the rest of the world is yet to catch up with.”

A Feed That Behaves Differently in Sarajevo

Across 46 countries surveyed on their news habits, interest in political news has fallen every year for the last decade, and trust in it declined in 16 of the 19 countries where trust itself was measured, a slide that hostile actors sometimes try to accelerate on purpose, as one case study of a pro-Russian hacking collective’s campaign against Finnish government institutions shows. Even people who go looking are mostly looking elsewhere: web-browsing records from large samples in the United States, Poland and the Netherlands found that only 3.4% of all the web addresses people visited were news sites at all, and more than half of what people read even on openly partisan sites was sport or recipes, not politics. On Facebook, news makes up less than 3% of the average American user’s feed, and on YouTube, the platform’s own recommendations account for over 70% of what gets watched, which sounds like exactly the algorithmic funnel critics of social media describe. But independent audits of that funnel keep finding it fairly boring: several changes tested on Facebook and Instagram, including removing shared posts and turning off content ranking altogether, produced no detectable shift in polarization, and YouTube’s own recommendations track what users already prefer more than they redirect it. The starkest illustration of how differently the same technology gets used comes from China, where the government has been shown to generate about 448 million fabricated social media posts a year, not to argue with critics online but to bury political conversation under a flood of distraction.

None of this proves social media is politically inert outside the US, only that the handful of rigorous causal tests run so far disagree with each other by geography rather than confirming a single global mechanism. The review’s authors caution that most of the comparative evidence, including the 53-country and 46-country surveys above, is observational, which can show that trust and polarization move together across countries without proving which one is pulling the other.

“Polarisation is not a single phenomenon with a universal set of causes or solutions,” says Andrea Baronchelli, a professor of complexity science at City St George’s University of London and one of the review’s senior authors. “Political institutions, social identities and media environments differ across countries, and the same intervention can have very different, even opposite, effects in different contexts.”

Since 2023, Meta, X and Reddit have all curtailed the kind of data access that let researchers run large comparative studies in the first place, and the European Union’s Digital Services Act, meant to force platforms back open, has so far seen patchy compliance, including a formal finding against X for failing to give researchers the access the law requires.

That scarcity bites hardest on the questions with the most at stake right now: a global survey covering the elections held in 2024 found that early fears about generative AI swinging votes were largely overblown, yet the tools built to police AI-generated content carry the same geographic blind spot as the polarization research itself, with Google’s own toxicity-detection system shown to flag ordinary German text as toxic far more readily than equivalent text in other languages tested. A test built and tuned mostly on American and English-language data, in other words, is not obviously the right test to run on a Turkish election or a German comment thread. The review’s own recommendation is unglamorous: fund and publish more research outside the United States, treat a paper’s national context as a fact worth stating rather than an afterthought, and stop assuming that whatever calmed a Facebook feed in Wisconsin will do the same in Sarajevo.

New Zealand’s polarization doubled the moment its electoral map changed, without a single social media platform involved, which is worth sitting with the next time an AI regulation gets drafted with only one country’s politics in mind. The technology may be global. The politics, this review argues, are not yet understood that way.

FAQ