In 2020, Facebook ran an experiment on its own users. For three months, thousands of people saw posts in simple chronological order instead of the usual algorithmic feed. The company wanted to know: would removing the algorithm reduce political polarization? The answer, published in Science three years later, surprised almost everyone. It didn't matter. People remained just as polarized, just as angry, just as divided.

This finding should have settled the debate about social media's role in our fractured politics. Instead, it opened a more complicated question: if algorithms aren't the root cause of polarization, why do they keep making it worse?

The Speed Problem

A study from Northeastern University published in 2025 offers the clearest evidence yet of algorithms' accelerating effect. Researchers led by Chenyan Jia created a browser extension that reranked X (Twitter) posts in real-time, either increasing or decreasing exposure to partisan animosity and antidemocratic attitudes. They didn't remove content or censor anyone—they simply changed what appeared first.

In one week, users' feelings toward the opposing political party shifted by about 2 points. That's the same change researchers typically observe over three years of normal life. The algorithm didn't create the polarization. It compressed time.

The effect held across the political spectrum. Liberals and conservatives both moved further from the center when algorithms fed them more partisan content. Over 1,200 participants confirmed what many suspected but few had proven: the feed doesn't just reflect our divisions. It speeds them up.

The Asymmetry of Engagement

Not all algorithms work the same way, and not all political content gets equal treatment. A 2023 field experiment on X, published in Nature, found that switching users from chronological to algorithmic feeds increased engagement and shifted political opinions toward more conservative positions. The reverse wasn't true—switching from algorithmic to chronological feeds produced no comparable effect.

X's algorithm appeared to promote conservative content while demoting posts from traditional media outlets. Users exposed to the algorithmic feed started following conservative political activist accounts, and they kept following them even after switching back to chronological order. The algorithm created habits that outlasted the algorithm itself.

Facebook showed similar patterns during the 2020 election. Studies found that sources favored by conservative audiences were more prevalent on the platform than liberal sources, and most misinformation sources were favored by conservative audiences. The average user received about half their content from people and pages sharing their beliefs—a level of ideological segregation far exceeding what researchers had found in internet browsing behavior.

These findings don't suggest a conspiracy. They reveal how engagement-maximizing algorithms interact with human psychology and existing political asymmetries. Conservative content may simply generate more of the reactions—shares, comments, angry emoji—that algorithms interpret as engagement.

The YouTube Exception

YouTube complicates the story. Research from the University of Pennsylvania's Computational Social Science Lab, published in 2024, found that YouTube's recommendation algorithm actually has a moderating effect. Researchers created bots trained on nearly 88,000 real users' watch histories and discovered that relying exclusively on YouTube's recommendations resulted in less partisan consumption than user-driven choices.

When users changed their viewing habits, sidebar recommendations shifted toward moderate content after about 30 videos. The algorithm "forgot" their previous preferences relatively quickly, at least in the sidebar. The homepage adjusted more slowly, catering to longer-term patterns.

This suggests the problem isn't algorithms per se, but specific design choices. YouTube's recommendation system apparently optimizes for watch time differently than Facebook and X optimize for engagement. Longer videos and sustained attention may favor different content than quick reactions and shares.

What Algorithms Actually Do

The evidence points toward a specific mechanism. Algorithms don't create political division—that started decades before social media, driven by party realignment, geographic sorting, and hyper-partisan cable news. But algorithms act as accelerants and amplifiers.

They compress the timeline of polarization, turning years into weeks. They create echo chambers more segregated than our offline communities. They reward content that triggers strong reactions, and in our current political environment, that often means partisan anger. They build habits that persist even when the algorithm changes.

A quarter of US adults now report social media as their primary news source. Half get news from these platforms at least sometimes. For these users, the algorithm isn't just one influence among many—it's the primary filter determining what political information they see.

Fifteen researchers writing in Science in 2020 captured the consensus: social media companies "have played an influential role in political discourse, intensifying political sectarianism." A Brookings Institution review of over 50 studies concluded that platforms "likely are not the root causes of political polarization, but they do exacerbate it."

Beyond the Binary

The debate over algorithms and polarization has suffered from binary thinking. Either algorithms cause polarization or they don't. Either we regulate them or we don't. Either users are victims or they have agency.

The research suggests a more nuanced reality. Users do seek out polarizing content—the YouTube study proves that left to their own devices, people often choose more extreme material than algorithms would recommend. But algorithms still matter because they determine the information environment where those choices happen. They decide what's easy to find and what requires effort. They shape what feels normal.

The Northeastern study's method points toward solutions. By reranking rather than removing content, researchers avoided censorship while still changing outcomes. Platforms could design algorithms that slow down polarization instead of speeding it up, that reward cross-partisan engagement instead of tribal signaling, that make moderation easier to find than extremism.

Mark Zuckerberg told Congress in 2021 that "it's not at all clear from the evidence or research" that social networks are polarizing. He was technically correct—they're not the main driver. But the evidence since then shows they're a powerful accelerator. In a country already divided, speed matters. What takes three years versus one week might mean the difference between a democracy that bends and one that breaks.