In November 2025, researchers at Northeastern University published a finding that should alarm anyone who's checked their phone today: they'd managed to shift people's partisan feelings by two points in a single week. That's the same movement typically observed over three years of normal political life. The culprit wasn't propaganda or persuasive arguments. It was simply adjusting which posts people saw on X, formerly Twitter.
The Experiment That Proved What We Suspected
Chenyan Jia and her team recruited over 1,200 people and divided them into groups during July and August 2024—a period that included Biden's withdrawal from the presidential race and the attempted assassination of Donald Trump. Using a browser extension powered by a large language model, they reranked posts in users' feeds without removing anything. The changes were imperceptible. Users thought they were seeing their normal timeline.
Some participants got more content expressing antidemocratic attitudes and partisan animosity. Others got less. Within a week, measurable shifts appeared in how both Republicans and Democrats felt about the opposing party. The algorithm didn't persuade anyone of specific policy positions. It simply changed the emotional temperature by controlling exposure.
What makes this particularly noteworthy is that the researchers didn't need Facebook or Twitter's cooperation. They built their own tool and made it open source, allowing other scientists to study algorithmic effects without waiting for tech companies to grant access to their black boxes.
When America Started Splitting
The timeline matters. Pew Research documented the steepest increase in political division between 2011 and 2014—exactly when Facebook, Instagram, Twitter, and YouTube hit peak user growth. In 1994, just 10% of Americans held consistently liberal or consistently conservative views across multiple issues. By 2014, that figure had doubled to 21%.
The partisan overlap that once characterized American politics essentially vanished. By 2014, 92% of Republicans positioned themselves to the right of the median Democrat, while 94% of Democrats sat to the left of the median Republican. More concerning, the share of each party holding highly negative views of the other more than doubled during this period.
Among the politically engaged, the shift was even sharper. Nearly 38% of active Democrats became consistent liberals by 2014, up from 8% two decades earlier. For engaged Republicans, consistent conservatives grew from 23% to 33%. These weren't people casually following politics. They were the ones most likely to be active on social media.
How Algorithms Choose What You See
Social media platforms don't optimize for truth or nuance. They optimize for engagement, which means they prioritize content that triggers emotional responses. Facebook's algorithm explicitly preferences posts that keep users scrolling. YouTube recommends videos based on watch history. Twitter suggests accounts to follow. Instagram curates an explore page. Each decision point nudges users toward more of what they've already seen.
The consequences play out in measurable ways. Fake news spreads on Twitter six times faster than legitimate news. A 2017 Japanese study found that even when Twitter users followed people with different viewpoints, they rarely discussed overlapping issues—the algorithm's design created echo chambers despite apparent ideological diversity.
When people spend time in homogeneous discussion groups, they don't moderate. They radicalize. Frequent interaction with like-minded peers pushes individuals toward more extreme positions. By 2014, people with down-the-line ideological views, particularly conservatives, were significantly more likely to report that most of their close friends shared their political perspectives.
The Radicalization Pipeline
Between 2014 and 2017, during a period when social media algorithms aggressively recommended content to maximize engagement, extremist groups like the Proud Boys, Generation Identitær, and QAnon all emerged and grew online. The platforms weren't intentionally promoting extremism, but their recommendation systems discovered that outrage and conspiracy theories kept people clicking.
Computer scientist Jaron Lanier has argued that these algorithms are designed to make users "hostile and paranoid" toward opposing viewpoints—not because tech companies want division, but because division drives engagement. He's suggested that even high-profile figures like Donald Trump may have been shaped by this dynamic, developing more extreme positions through Twitter addiction and the feedback loop it created.
The platforms have tried to respond. Facebook stopped recommending political groups to users in 2021. Twitter and Facebook deployed enhanced machine learning during the 2020 election to limit misinformation spread. But extremists adapt, using cryptic language and subtle signals to circumvent content moderation.
The Paradox of Exposure
Here's where the problem gets thornier: simply exposing people to opposing viewpoints doesn't fix polarization. Research shows that once group polarization affects someone, they tend to regard opposing perspectives as attacks on their identity. Exposure doesn't moderate their views—it affirms their negative attitudes toward political opponents.
This helps explain why Republicans now trust fewer news sources than they did a decade ago, with Fox News, Trump's speeches, and social media posts among the few they regularly consume and believe. Divergent interpretations of events get fed back through social media, creating a feedback loop where each side becomes more convinced the other is delusional or dangerous.
Most intense partisans now believe the opposing party's policies genuinely threaten the nation's well-being. That's not a policy disagreement. It's an existential threat assessment.
Redesigning for Democracy
The Northeastern study suggests a path forward, though not an easy one. If algorithms can shift partisan feelings by two points in a week, they could theoretically shift them in the opposite direction. The researchers hope platforms might redirect their systems away from pure engagement metrics toward "positive societal impact, like reducing affective polarization."
The challenge is structural. Social media companies make money by keeping users engaged, and engagement correlates strongly with emotional intensity. Asking them to voluntarily reduce polarization is asking them to potentially reduce profits. Regulation could mandate algorithmic transparency or require platforms to test for polarizing effects, but that faces both technical and political obstacles.
What the research makes clear is that this isn't a problem of individual choice or media literacy. When a week of algorithmic adjustment produces three years' worth of partisan shift, we're dealing with an environmental factor—like lead in gasoline or smoking in restaurants. The question isn't whether people should be more rational. It's whether we'll redesign the systems shaping how millions encounter political information every day.