Tribalism and Polarization

In the autumn of 2024, during the final weeks of a bitter US presidential campaign, 1,256 people agreed to let researchers quietly rewire what they saw online. Each installed a browser extension on X, the platform formerly called Twitter. The extension did something no ordinary user could do: as posts loaded into the feed, a large language model read each one, scored it for how much hostility it expressed toward the political opposition, and reshuffled the results—pushing the most antagonistic material down and letting calmer content rise. Nobody was blocked from anything. No accounts were banned. The only change was the order in which posts appeared.

Ten days later, the researchers measured how participants felt about the other party using a standard "feeling thermometer," a zero-to-one-hundred scale political scientists have used for half a century. The people whose feeds had been de-escalated felt warmer toward their political opponents by roughly two points. That may sound trivial until you learn the baseline for comparison: across the American electorate, feelings toward the opposing party have historically cooled by about that same two points over three years. A week of reordered posts had moved the needle as far as three years of ordinary life—in the opposite direction.

The study, published in Science in November 2025 under the title "Reranking partisan animosity in algorithmic feeds," is the closest thing we have to a controlled experiment on one of the defining questions of the age: does the machinery that curates our information actually make us hate each other more, or does it merely reflect hatred that was already there? For years that question was fought over with correlations and competing surveys. This experiment cut through much of the fog. It did not simply observe that hostile people use social media. It intervened—changed one variable, the ranking algorithm—and watched attitudes move. The direction of causation was no longer a matter of interpretation.

What the experiment can and cannot tell us

It is worth being precise about the magnitude, because the number is easy to misread in both directions. Two points on a hundred-point thermometer is not a personality transplant. Nobody in the study crossed the aisle. What changed was a small, measurable shift in warmth toward the outgroup, produced by a one-week nudge to a single platform's feed. The significance lies not in the size of the effect but in the ratio between effort and result: a modest algorithmic tweak, applied for days, produced a change comparable to what unfolds over years of accumulated political experience. If a week of gentle de-escalation moves attitudes that much toward tolerance, the same lever, pulled the other way for years, plausibly moves them a great deal toward animosity.

That inference—running the logic in reverse—is where the study becomes something more than a hopeful footnote. The engagement-optimized feed that billions of people scroll every day is, in effect, the untreated condition. It ranks content not for civility but for reaction, and the reactions it rewards most reliably are anger and tribal solidarity. The experiment did not need to prove that ordinary feeds polarize; it demonstrated that reversing their tendency de-polarizes, quickly and measurably, which makes the polarizing tendency of the default very hard to explain away.

The researchers found something else worth dwelling on. Downranking hostile content didn't only cool participants' feelings toward opponents—it lowered their self-reported anger and sadness. The emotional weather of political media, in other words, is substantially a product of what the algorithm chose to show. People who feel worn down and furious after an hour of scrolling are not simply reacting to a grim world. They are reacting to a grimness the ranking system selected and amplified on their behalf.

Why the machine tilts this way

None of this happens because engineers set out to divide people. It happens because of what they were optimizing for.

Recommendation systems are trained to maximize engagement—time on platform, clicks, shares, replies, the small behavioral signals that translate into advertising revenue. Engagement is the target; everything else is a side effect. And engagement, study after study has found, runs highest for content that provokes strong emotion, especially the hot emotions: outrage, indignation, tribal triumph. Calm, nuanced, cross-cutting content simply does not generate the same reaction, so the system learns to show less of it. The result is a machine that, without any instruction to do so, discovers that division is engaging and acts accordingly.

The mechanism is a feedback loop that tightens with every interaction:

graph LR
  A[User engages with content] --> B[Algorithm infers preferences]
  B --> C[Surfaces similar, high-reaction content]
  C --> D[User engages more with confirming, tribal posts]
  D --> E[Information diet narrows and homogenizes]
  E --> A

Each turn of the loop nudges the feed toward material that confirms what the user already believes and portrays the other side as threatening, because that material holds attention best. Over months, the diet becomes ideologically uniform—not because anyone chose uniformity, but because uniformity is what maximizing reaction produces. This is why the polarizing effect should be understood as an emergent property of the incentive structure rather than a design decision. You could not remove it by finding and firing the person who built the "make people angry" feature. There is no such person and no such feature. There is only a system rewarded for engagement, discovering on its own that engagement and animosity travel together.

A striking piece of corroboration comes from simulation. Researchers have built artificial social media environments populated by AI agents and swapped out the underlying language model that drives recommendations. Regardless of which model they used, the platforms reliably developed echo chambers, concentrated influence in a few loud voices, and amplified extreme content. When the same pathology appears no matter which engine you install, you are not looking at a bug in one company's code. You are looking at a property of the objective itself.

What the narrowing does to judgment

The most consequential damage from a homogenized feed is not that people see more of what they already believe. It is what happens to the mental muscle that lets us take an opponent seriously.

Charitable evaluation of an opposing argument—the discipline of assuming the other person might have a point, and testing your own view against theirs—is a skill, and like any skill it decays without practice. In a diverse information environment, that practice is unavoidable; you keep bumping into people who disagree and have to reckon with them. In an algorithmically sorted one, the practice disappears. Disagreement stops registering as a legitimate difference of values and starts feeling like ignorance or bad faith. Why would anyone believe that, unless they were stupid or lying?

Compounding the problem, the opposing side you do encounter is systematically the worst version of it. Extreme content generates the most engagement, so extreme voices surface most often. The reasonable conservative or thoughtful progressive who makes up the bulk of the other party rarely goes viral; the inflammatory outlier does. Over time, a user's entire mental model of "the other side" is assembled from its most enraging exhibits. You are not shown the median opponent. You are shown the caricature, again and again, until the caricature becomes the face.

Networked tribalism

Filter bubbles and echo chambers describe what an individual sees. Their social consequence, playing out across whole communities, has acquired its own name: networked tribalism—the digital tribes that form when filter bubbles, engagement algorithms, and political polarization reinforce one another. It is not the same thing as having a political opinion. It is a mode of belonging, and it rests on four dynamics that work together.

The first is intensified in-group identity: continuous immersion in a like-minded community makes the group central to how members see themselves. The second is outgroup hostility, which is not incidental to the first but structural to it—the more cohesive and cherished the in-group, the more sharply the out-group must be drawn as a threat to sustain that sense of belonging. The third is conformity pressure: within the tribe, expressing doubt or nuance risks social rejection, so members self-censor deviations and the visible consensus grows more uniform and more extreme than the private views underneath it. The fourth, and most corrosive, is the conversion of truth into tribal narrative: when outside information contradicts the group's story, members face a social incentive to reject the information rather than revise the story, because the story is load-bearing for their identity. Facts become negotiable in a way loyalty is not.

Put these four together and you get a ratchet. Signaling commitment requires expressing outrage; expressing outrage raises the temperature; the raised temperature demands still louder signals. Emotional escalation becomes the dominant form of participation, and it only turns one way. The unsettling finding from both experimental and longitudinal work is that this happens to people who consider themselves reasonable and open-minded. The transformation is gradual and largely invisible from the inside. Many who go through it only recognize it in hindsight—when a once-friendly relationship with a relative who votes differently has curdled, and they cannot quite say when it happened.

The measurement debate

If algorithmic polarization is so powerful, why do serious researchers still argue about how common echo chambers are? The answer is a genuine and instructive methodological split.

Studies built on computational methods and behavioral data—who follows whom, which posts spread through which networks, how exposure patterns differ across communities—tend to support the echo chamber hypothesis robustly. They see strong ideological sorting in the actual traffic of the platform. Studies built on surveys and self-reported exposure often find the opposite: that plenty of users do encounter cross-cutting content, that their information diets are less sealed than the echo chamber picture implies.

The two camps are not really contradicting each other so much as measuring different things. There are three distinct phenomena here—what users see, what they engage with, and what affects them—and they do not coincide. A user may scroll past an opposing viewpoint (it appears in the "see" column), decline to click or share it (absent from "engage"), and remain wholly unmoved by it (absent from "affect"). Survey methods, which lean on what people recall seeing, capture the first and widest column. Behavioral methods capture engagement. The 2025 Science experiment reached the third and most important one—effect—by manipulating the feed and measuring attitude change directly. That is why it partly dissolved the debate rather than joining it. Whether or not strict informational isolation is common, engagement-optimized curation demonstrably shifts attitudes, and altering the algorithm alters the attitudes within days. The practical stakes do not hinge on winning the prevalence argument.

The persuasion machine

A separate development sharpens the picture considerably. Alongside recommendation feeds, we now have AI chatbots built explicitly to change political minds—and they work.

Experiments conducted across the United States, Canada, and Poland in 2024 and 2025 found that short conversations with a well-designed persuasive AI moved voters' stated preferences by several points on a hundred-point scale, using calm, evidence-based dialogue rather than manipulation. That may not sound dramatic, but consider what it combines. Interpersonal persuasion has always been the most effective form—a patient one-on-one conversation, tailored in real time to the other person's specific worries, beats any billboard or broadcast. Its fatal limitation was that it did not scale; you can only have so many conversations. Automation removes that ceiling. A conversational AI can hold a million individually tailored dialogues at once, each adapting its arguments to the particular concerns, emotional triggers, and reasoning style of the person on the other end.

This is qualitatively different from the political targeting that preceded it. The Cambridge Analytica era relied on crude demographic buckets—show this ad to suburban women aged 35 to 50. A persuasive chatbot operates at the granularity of the individual, in real time, in dialogue, learning as it goes. It brings the intimacy of one-on-one persuasion to the scale of mass media, a combination that simply did not exist before.

The dual use

The same capability points in two opposite directions, and this is the crux of what makes it hard to govern.

Deployed constructively, conversational AI could do something neither mass media nor human conversation can manage: expose people to well-reasoned alternative perspectives, patiently and at scale, cooling polarization across millions of interactions. The 2025 reranking study already showed that de-escalation is technically achievable; a persuasion engine aimed at bridging rather than dividing could push in the same direction far more actively. Deployed exploitatively, the identical technology becomes an instrument of hyper-personalized radicalization—AI agents that map an individual's psychological profile and then deliver precisely calibrated messaging to harden bias or drag them toward the extremes, one tailored conversation at a time, at a scale of millions.

Neither scenario is hypothetical; the technology for both already exists. What determines which use predominates is not the technology but the surrounding conditions—regulatory frameworks, platform incentives, and the priorities of whoever builds and deploys these systems. And at present those conditions tilt the wrong way. Commercial incentives favor persuasion over depolarization, and the rules governing AI in political communication remain thin relative to how fast the capability is advancing. The dual-use character is not a comforting "it could go either way." It is a warning that, absent deliberate choices, it will go the profitable way.

What it does to a person

Polarization is usually discussed as a political problem. It is also a health problem, and the Science experiment gave us a clean measurement of it: reducing exposure to hostile content lowered participants' anger and sadness. Read that in reverse and it describes the standing condition of heavy users—an elevated, chronic negative emotional load that people experience as merely "keeping up with the news" but which is substantially manufactured by what the algorithm selects.

The harms operate through several distinct mechanisms that reinforce one another. Chronic stress activation comes first: a feed that frames the other side as an existential threat keeps the body's stress response switched on, and the cumulative toll on mental and physical health resembles that of other chronic stressors. Relationship damage follows as tribal identity begins to override personal ties—not because political disagreement is new, but because the intensity and uniformity of algorithmically reinforced identity makes the ordinary accommodation of difference harder to sustain, so friendships and family dinners fracture along lines that once bent. Reduced empathetic capacity is a third: sustained portrayal of opponents as malicious does not destroy the ability to empathize so much as make empathy feel factually unwarranted—why extend understanding to people your information environment insists are acting in bad faith? Fourth is cognitive rigidity, the atrophy of the capacity to consider alternatives and update on new evidence, a skill that withers without the exercise a diverse diet would provide. And fifth, most consequential, is a distortion of proportion: when tribal political identity becomes central to the self, ordinary policy disagreements start registering as existential threats, producing emotional responses wildly out of scale with any realistic assessment of the stakes.

These effects compound quietly, which is exactly what makes them hard to catch. The gradualism is not incidental to the process; it is part of what makes it work. People rarely notice the day they became angrier and more certain than their own prior values would have predicted.

The risk to democracy

Scale these individual changes up to a population and you reach the reason this chapter belongs in a book about the future of societies rather than a manual on personal well-being.

Democratic systems are not self-sustaining machines. They run on a handful of conditions that extreme polarization actively corrodes. The first is a shared factual baseline—enough common ground about what is true to make debate possible at all. When opposing communities inhabit separate informational realities, curated by separate feeds, there is no longer a shared field on which to disagree; argument gives way to mutual incomprehension. The second is a willingness to treat opponents as legitimate rather than existential. When the other party is experienced not as fellow citizens with different values but as a threat to national survival, compromise stops looking like governance and starts looking like betrayal, and the ordinary machinery of give-and-take seizes up. The third is acceptance of the legitimacy of outcomes—the peaceful transfer of power when your side loses. When electoral defeats are reflexively read as fraud rather than as genuine results, the procedural norms that let democracy function become contingent on winning, which is to say they stop being norms at all.

Research on polarization consistently finds it associated with political violence, democratic backsliding, erosion of civil liberties, collapsing institutional trust, and social fragmentation. Here honesty requires care about the strength of the claim, and it cuts in a specific direction. These are robust associations, documented across many countries over recent decades—but they are associations. The causal chain running from a specific algorithm to a specific instance of democratic backsliding is not something any single study has cleanly established, and it would be overreaching to assert it. What the evidence supports is a cumulative argument rather than a single proven arrow: engagement-optimized curation demonstrably intensifies polarization (the experiment); polarization is reliably associated with democratic dysfunction (the observational literature). The bridge between them is strong enough to take seriously and warrant action, but it is a bridge built of two spans, and the joint deserves to be named rather than papered over. The stronger version—"AI feeds are dismantling democracy"—outruns the evidence. The defensible version—"AI-driven polarization plausibly erodes conditions democracy needs, and we have good reason to treat that as urgent"—is what the data will bear.

Where in the arc are we?

Is this trend accelerating, stabilizing, or already turning? The honest answer is that we are still on the way up, with the first faint evidence that intervention is possible.

The forces driving polarization remain fully intact. The business model still rewards it, the persuasion tools are getting more capable, and no major platform has voluntarily reoriented its core ranking system away from engagement. Against that, the 2025 Science study is the clearest sign yet that the trajectory is not fixed—a demonstration that the same lever driving polarization can be pulled the other way, quickly. But a research browser extension used by a thousand consenting participants is not a deployed intervention, and there is no sign the platforms intend to adopt anything like it at scale. A meaningful turning point would look like something specific: a major platform changing its default ranking to de-emphasize animosity and keeping it changed despite the engagement cost; or a regulatory regime that forces the incentive to shift regardless of what platforms prefer. Neither has happened. We have proof the disease responds to treatment and no evidence the treatment is being administered.

The clock and the rulebook

The pace is the whole problem. If it took a generation for algorithmic curation to shape political attitudes, institutions might adapt in time, the way they eventually adapted to broadcast television. But the Science experiment measured attitude change in days. When the technology reshapes political psychology on a timescale of a week and governance frameworks develop on a timescale of years, the two are not merely mismatched—they are in different eras. Every month the rulebook lags, the capability compounds, and some fraction of the resulting attitudinal change may prove sticky in ways we cannot yet reverse. That is the argument for urgency: not certainty about irreversible damage, but a rate of change fast enough that waiting to be certain is itself a decision with consequences.

Who is responsible, and what could change

This brings us to the questions the evidence forces but cannot settle on its own.

Do platforms bear responsibility for the polarization their algorithms produce? The strongest version of the argument does not require proving intent. It runs like this: the harm is predictable—the research showing that engagement optimization amplifies animosity is now robust and public. The design is profitable—polarized users are engaged users, spending more time, generating more data and more ad revenue. And the harm falls on a public good, democratic functioning, for which the platforms bear no accountability under current law. When a company profits from a design it can foresee will damage a shared public good, that is a recognizable category of harm, and the case that regulatory frameworks should address it does not depend on anyone at the company wishing for the damage. The counterargument—that platforms merely reflect what users choose, and that users prefer homogeneous, confirming content—has real force, but it is weakened precisely by the reranking result: change the feed and attitudes change, which means the platform is not a neutral mirror but an active shaper of the outcome.

If reform is warranted, which of the available interventions could actually happen? Five are usually proposed, and they differ sharply in feasibility once you take the business model into account.

The uncomfortable through-line is that the interventions best aligned with democratic health are worst aligned with platform revenue. Algorithmic redesign is the most direct fix and the least likely to be adopted voluntarily, because it asks a company to lower the metric its valuation depends on in exchange for a social good it is not accountable for producing. Voluntary reform becomes viable only under a narrow condition: when the reputational or user-trust cost of polarization starts to outweigh its engagement value—when staying divisive becomes commercially expensive. Absent that, the realistic path to reform runs through regulation, which reshapes the incentive from outside rather than waiting for it to change from within. Of the five, the two with the best odds are the ones that do not depend on platforms wanting to change: binding regulation, and the sort of sustained public pressure that shifts the commercial calculus. The rest are feasible only after the incentive has already moved.

And persuasive AI specifically may warrant its own category of rules. The general-purpose frameworks now being drafted for AI treat it as one technology among many, but a system engineered to change political minds at individual granularity and mass scale poses risks that a medical-diagnosis AI or a coding assistant does not. Proportionate rules would focus on three things: disclosure, so that people know when they are being persuaded by a machine rather than a person; targeting limits, restricting the psychological micro-tailoring that turns persuasion into manipulation; and electoral-context constraints, tightening both around campaigns, when the stakes for democratic legitimacy are highest. The aim is not to ban political speech by AI but to prevent the covert, hyper-personalized version that the technology now makes possible.

What we still do not know

It would be dishonest to end without marking the edges of the evidence, because the most important questions remain open.

The 2025 Science study, for all its power, is a single ten-day experiment with 1,256 participants on one platform during one country's presidential campaign—a moment of peak political intensity that may not generalize to ordinary times or other cultures. We do not know whether its effects hold on TikTok, YouTube, or Facebook, whose formats and audiences differ; whether they hold outside an election's heat; or whether they hold across the world's varied political systems. Establishing durability would require multi-platform, multi-country replication over far longer windows than ten days.

Nor do we know whether the attitude changes are durable once the exposure stops. The experiment measured a shift while the intervention was running. It could not tell us whether a de-escalated week leaves any lasting trace after the extension is uninstalled, or whether attitudes snap back—and by the same token, whether the polarizing effects of ordinary feeds persist after someone logs off or fade quickly. That is a question only a longitudinal study can answer: follow the same people for months or years, measure them repeatedly, and see what sticks. Until such work exists, we cannot say whether the political psychology reshaped by algorithms is a temporary state that lifts when the feed changes, or a durable rewiring that outlasts it. That uncertainty is not a reason to dismiss the concern. It is a reason to fund the research and to act, in the meantime, as though a fast, powerful, and possibly durable effect on the foundations of democratic life is exactly what it appears to be.

Summary

1. The landmark finding. A November 2025 Science experiment ("Reranking partisan animosity in algorithmic feeds") had 1,256 people run a browser extension that used a large language model to downrank hostile content on X for ten days during the 2024 US campaign. In one week, feelings toward the opposing party warmed by about two points—a shift that takes roughly three years to occur naturally. Run the logic in reverse, and years of engagement-optimized feeds plausibly polarize by a great deal.

2. It's the incentive, not a decision. Recommendation systems maximize engagement; engagement is highest for outrage and tribal solidarity; so the systems amplify division as an emergent property of the objective, not a deliberate design. Simulations confirm this appears regardless of which underlying model is used.

3. Narrowing erodes judgment. A homogenized feed lets the muscle of charitable evaluation atrophy and, because extreme content travels furthest, makes the caricature of the other side its representative face.

4. Networked tribalism is the community-level result, composed of in-group identity, outgroup hostility, conformity pressure against dissent, and the conversion of truth into tribal narrative—a ratchet that only escalates.

5. Persuasive AI combines the effectiveness of one-on-one dialogue with the scale of automation, making it qualitatively different from crude demographic targeting. It is dual-use: the same tool could depolarize at scale or radicalize with hyper-personalization. Current incentives favor the exploitative use.

6. Personal harm accumulates through chronic stress, relationship damage, reduced empathy, cognitive rigidity, and out-of-proportion emotional responses—gradually and largely unnoticed.

7. Democratic risk comes from eroding shared facts, framing opponents as existential threats, and questioning electoral legitimacy. The evidence supports the defensible version of this claim (a strong two-span argument from experiment plus association), not the overreaching version of a single proven causal arrow.

8. The clock outpaces the rulebook. Attitudes move in days; governance moves in years. That mismatch, not proven irreversibility, is the case for urgency.

9. Responsibility and reform. Profitable, predictably harmful design aimed at a public good is a recognizable category of harm. Of five interventions—algorithmic redesign, user control, transparency, nudges, regulation—the two least dependent on platform goodwill (binding regulation and cost-shifting public pressure) are the most feasible, because the fixes best for democracy are worst for revenue. Persuasive AI may warrant its own rules on disclosure, targeting, and electoral use.

10. What we don't know. One ten-day, single-platform, single-election study cannot establish durability, cross-platform generality, or whether the effects persist after exposure ends. Longitudinal, multi-platform research is needed—and its absence is a reason to act, not to wait.

Sources

• Reranking partisan animosity in algorithmic feeds | Science
• Social media research tool can reduce polarization | University of Washington
• Social media tool lowers political temperature | Stanford Report
• AI tool decreased political polarization | Scientific American
• Don't blame the algorithm: Polarization inherent in social media | Science
• How Does Social Media Impact Political Polarization? | Northeastern University
• AI and Social Media: A Political Economy Perspective | MIT
• Networked Tribalism: Filter Bubbles + AI Algorithms | Voices of VR
• AI Echo Chambers and Democracy | TechRxiv

Last updated: 2026-09-04