Algorithms Shift Polarization. Why Does Policy Still Miss the Real Problem?

By Dr. Matthias J. Becker

AddressHate Research Scholar at NYU’s Center for the Study of Antisemitism | Lead, Decoding Antisemitism | Founding Research Advisor, AddressHate

Image: Google DeepMind

A landmark study provides some of the clearest causal evidence to date that platform architecture drives political hostility. The policy response reveals how wide the gap remains between what research shows and what gets addressed.

---

Introduction

A breakthrough study published in Science this November provided some of the clearest causal evidence to date that social media algorithms shape political polarization. Researchers at Stanford built a browser extension that independently reranked users’ feeds in real time, demonstrating that algorithmic choices directly altered users’ political attitudes by amounts comparable to several years’ worth of polarization change in long-term US surveys.

For researchers who study online radicalization, the findings confirmed what extensive observational evidence had long suggested: platform architecture—what gets amplified, how content is ranked, which posts surface first—directly shapes democratic outcomes. The study was methodologically rigorous, causally clean, and measured effects that matter.

For the first time, independent researchers demonstrated algorithmic causation without requiring access to platform infrastructure—a milestone in platform accountability research.

Yet the policy conversation remains largely elsewhere.

Policy attention has focused instead on symbolic fixes, geopolitical narratives, and interventions that are politically legible, easier to implement, and do not require confronting platform business incentives. Around the same time the Stanford study appeared, lawmakers from both parties welcomed X’s new location feature, which displays the country from which an account operates, framing it as a step toward combating foreign influence and online antisemitism.

The Stanford study should reshape the conversation about platform accountability. Instead, the policy response that followed reflects a familiar pattern: research identifies mechanisms; policy focuses on symptoms.

This pattern repeats across digital governance. Breakthrough research isolates the forces shaping democratic outcomes. Policy attention gravitates toward solutions that fit existing narratives about external threats, require minimal structural change, and avoid examination of the architectures that enable radicalization.

The result is a widening gap between what we know works and what we’re willing to demand.

What the Research Actually Shows

The Stanford study, led by researchers Tiziano Piccardi, Martin Saveski, and Chenyan Jia, represents both a methodological and substantive breakthrough. Working without platform permission, they built infrastructure to intercept and rerank X feeds in real time, using large language models to identify content expressing “antidemocratic attitudes and partisan animosity” (AAPA)—including partisan hostility, support for undemocratic practices, opposition to bipartisanship, and biased evaluation of facts.

Over 1,256 participants received either increased or decreased exposure to AAPA content for one week during the 2024 presidential campaign. The effects were measured through surveys embedded directly in users’ feeds and post-experiment assessments.

Reducing AAPA exposure led participants to feel warmer toward the opposing party by 2.11 degrees on a 100-point scale. Increasing exposure caused a symmetrical cooling of 2.48 degrees. Critically, Democrats and Republicans responded identically to the interventions—the effects were bipartisan. These are not huge shifts for any single individual, but applied to millions of users over months or years, they translate into substantial structural tilts toward hostility.

Moreover, 74% of participants reported noticing no impact on their experience. Algorithmic effects operate below conscious awareness, shaping attitudes without users recognizing the intervention.

The study demonstrates what years of correlational research could not: that algorithmic ranking decisions directly shape political attitudes. The content remained the same. The users remained the same. Only the ranking changed—and polarization moved accordingly.

Three Drivers—And Why Only One Gets Policy Attention

To understand why this evidence matters—and why policy responses miss the mechanism—we need a framework for how antisemitism and antidemocratic hostility actually spread online.

My research, analyzing over 300,000 web user comments across major news channels using qualitative discourse analysis supported by computational methods, reveals three distinct but interconnected drivers:

• First: Malicious actors. Foreign and domestic campaigns deliberately inject polarizing content. State-backed influence operations and coordinated disinformation networks exist. This is the driver dominating political rhetoric and policy attention.

• Second: Algorithmic amplification. Platform ranking systems, optimized for engagement, disproportionately surface emotionally intense content. Outrage drives clicks. Hostility generates shares. Algorithms don’t evaluate content quality or social consequences—they optimize for metrics that happen to correlate with polarization. The Stanford study provides causal evidence for this mechanism.

• Third: Communication conditions. The socio-technical environment of online spaces—such as anonymity that removes social accountability, mutual reinforcement that validates extreme positions, and omnipresent hate speech that normalizes extremity—fundamentally shapes how users behave and what content they produce.

The Stanford study measured exposure effects but did not examine how these conditions radicalize the ordinary users who create the content algorithms then amplify.

These three drivers interact dynamically. Malicious actors inject content. Algorithms amplify it. Communication conditions radicalize ordinary users who then produce more extreme content—which algorithms amplify further, reaching new users who become radicalized under the same communication conditions.

The Stanford study isolates the second driver and proves it matters causally. But understanding all three is essential because policy interventions typically address only the first—and often in ways that miss even that mechanism.

Why Geographic Solutions Miss All Three Mechanisms

The X location feature exemplifies how policy attention focuses on one driver while ignoring the others entirely.

The feature addresses exclusively malicious actors, specifically foreign ones. It assumes that knowing where an account is located helps identify bad actors and discount their content accordingly.

This assumption is flawed on multiple levels. Geographic origin tells us nothing about intent, coordination, or impact. An account in Eastern Europe may be an ordinary person with opinions about American politics. An account in Iowa may be part of a domestic influence operation. Location and malicious intent are orthogonal categories.

More fundamentally, focusing on foreign accounts ignores where most high-engagement polarizing content actually originates. In our analysis, antisemitic discourse surged to 36-38% of comments on major UK news outlet YouTube channels following the October 7 attacks. After the antisemitic hate crime in Washington in May 2025, such content averaged 43% across major English-language news channels, reaching 66% on some outlets.

While we cannot completely exclude some automated activity, the idiom, references, and interaction patterns overwhelmingly indicate domestic participation. High-profile domestic actors—politicians, media figures, influencers—routinely traffic in strategically ambiguous statements that seed problematic interpretations. Digital intermediaries sharpen and amplify these messages. Comment sections collapse the ambiguity into explicit hate speech.

Foreign interference exists and deserves scrutiny—but in the datasets we examined, domestic dynamics overwhelmingly drove the volume and intensity of antisemitic expression.

The implication is clear: even if the location feature perfectly identified every foreign account, it would leave the dominant source of polarizing content completely untouched.

What the Stanford Study Reveals—And What It Doesn’t

The Stanford research provides strong causal evidence for algorithmic amplification. The experiment demonstrated that changing how content is ranked changes political attitudes, even when the content itself remains constant. This is precisely the kind of mechanistic evidence that should drive policy.

But the study also reveals, by design, what remains unexamined: communication conditions.

The researchers showed that exposure to AAPA content increases polarization. They measured what happens when algorithms amplify certain material. What they could not measure was how and why that content emerged in the first place.

The AAPA posts that participants encountered were produced by users operating under specific communication conditions. Some content likely reflected what users would express in any context. But much was probably shaped by anonymity (enabling transgressive expression), mutual reinforcement (validating extreme positions), and normalized hostility (making extremity seem ordinary).

Communication conditions produce radicalized content. Algorithms then amplify it. This reaches users who are further radicalized by those same conditions, producing more extreme content in a self-reinforcing cycle.

The Science study isolates one mechanism in a much larger system. It shows, causally, that algorithmic amplification shapes political attitudes even when the underlying content and users remain constant. But amplification does not operate in isolation: it interacts with malicious actors who inject polarizing material and with communication conditions that shape what ordinary users express.

Location labels are framed as a response to malicious actors, yet they capture only the purported geography of an account—a poor proxy for the coordinated behaviors that actually matter. And crucially, they do nothing to address what the Science study demonstrates: how ranking systems elevate emotionally intense content and thereby shift attitudes at scale. In short, the policy response targets the most symbolically convenient driver while leaving the empirically important one untouched.

What Evidence-Based Accountability Would Require

If policymakers took the Stanford findings seriously, they would demand fundamentally different interventions.

For algorithmic amplification:

• Mandatory transparency into recommendation systems and ranking algorithms

• Independent researcher access without platform gatekeeping

• Regulation of engagement-optimization that profits from polarization

• Required testing of alternative ranking approaches, prioritizing democratic values

The Stanford researchers had to build their own infrastructure to study algorithmic effects. This shouldn’t be necessary. Platforms should be required to enable independent auditing of core ranking systems.

For communication conditions:

• Foundational: fine-grained multimodal discourse analysis.
  Before diagnosing communication conditions, we need a detailed empirical understanding of the discourse itself. This requires systematic, multimodal analysis of online hostility—its lexicon, metaphors, frames, narrative arcs, escalation patterns, and audience uptake over time. Without understanding how hate is communicated, we cannot meaningfully intervene in the conditions that enable it.

• Link communicative patterns to platform affordances.
  These patterns must then be connected to specific design features—ranking systems, recommendation engines, visibility mechanics, reply structures, quote-tweets, stitching, duets, and comment threading—that amplify conflict and facilitate in-group reinforcement. Affordances are not neutral; they shape discursive behavior.

• Design interventions targeting mutual-reinforcement loops.
  Once escalation points are empirically mapped, platforms can introduce interventions that disrupt the loops through which hostility is validated and intensified—dogpiling dynamics, outrage cascades, quote-wars, and recursive comment spirals.

• Address how anonymity shapes hostile discourse.
  Debates about anonymity should be grounded in observed communicative practices, not assumptions. Research consistently shows divergent effects: pseudonymity can enable performative transgression, whereas stable identity cues introduce reputational constraints.

• Introduce friction mechanisms along radicalization pathways.
  Deploy targeted friction—delays, prompts, contextual warnings, confirmation steps—at moments where discourse transitions from disagreement to dehumanization or annihilationist rhetoric. The aim is not censorship but interruption of momentum.

• Develop systems that surface normalization of extremity.
  Create user-facing indicators that reveal when conversations drift into patterns associated with dehumanization, conspiracism, or incitement. These signals translate discourse analysis into real-time literacy tools for users.

Without intervening at the level of communicative dynamics, platforms will continue to produce the very content that amplification systems learn to elevate.

For malicious actors:

• Detection focused on behavioral patterns, not geography

• Recognition that domestic actors often matter more than foreign ones

This requires rethinking fundamental platform architecture. But without addressing communication conditions, we leave intact the environment that continuously produces content for algorithms to amplify.

Why the Gap Persists

Why does policy gravitate toward interventions like the location feature rather than what evidence suggests would work?

The answer reveals the political economy of platform governance. The location feature requires almost nothing from platforms: add a label, done. No change to business models. No transparency into proprietary systems. No examination of how design enables radicalization.

Real accountability would require changes that platforms resist because:

• Transparency into recommendation systems threatens competitive advantage and reveals engagement-optimization mechanics

• Independent researcher access creates legal and PR risks

• Regulating engagement optimization could reduce user time and advertising revenue

• Redesigning communication affordances requires significant engineering resources

The contrast between what is easy for platforms and what is necessary for democracy explains the persistent gulf between research and regulation.

Some lawmakers may genuinely believe location labels increase transparency. But in practice, these measures function as accountability theater—the appearance of action without institutional change.

This dynamic is bipartisan. Republicans can condemn foreign influence without examining how right-wing media ecosystems mainstream conspiracy theories. Democrats can criticize online hate without confronting how some progressive rhetoric traffics in problematic tropes. Both parties position themselves as defenders against external threats while avoiding harder questions about domestic discourse, platform business models, and communication architecture itself.

Closing the Gap

The celebration of X’s location feature, in the same week that breakthrough research demonstrated what actually drives polarization, encapsulates the persistent gap between what research shows and what policy addresses.

We now have causal evidence that algorithmic amplification matters. We have extensive empirical evidence that communication conditions—anonymity, mutual reinforcement, omnipresent hate speech—radicalize ordinary users who produce the content algorithms amplify. And we have clear data showing that domestic actors and participatory dynamics drive most high-engagement polarizing content.

Against this evidence, focusing on account geography represents more than a missed opportunity. It represents a fundamental misunderstanding of how online radicalization operates.

Online antisemitism and antidemocratic hostility are not primarily imported threats. They are produced through the interaction of algorithms that reward polarization, communication conditions that enable radicalization, and malicious actors—often domestic—who exploit these dynamics.

If we’re serious about addressing online radicalization, we need interventions aimed at mechanisms, not geography. We need transparency into algorithmic systems. We need research into how platform affordances shape radicalization. We need policies addressing all three drivers, not just the politically convenient one.

The location feature tells us where an account is. The research tells us what actually matters: what content gets amplified, and how communication conditions shape radicalization. One is trivial to implement and largely ineffective. The other is difficult to address and essential to democracy.

The Stanford study demonstrated what becomes possible when research operates independently of platform gatekeeping. The policy response demonstrates how far we remain from demanding the accountability that evidence requires.

The evidence is clear. The question is whether our policies will finally catch up.