A new UC Berkeley Economics study finds that online news algorithms can quietly push readers toward more polarized views by reinforcing what they already believe.
The research from Berkeley Economics PhD Student Mingduo Zhao shows that even small differences in opinion can be amplified over time as algorithms learn what users click on and serve them more of the similar content. The result is a feedback loop that can deepen divisions, while also keeping users more engaged.
In his dissertation, paper titled, “News Consumption, Recommender Systems, and Polarization,” Zhao examined how recommender systems interact with user behavior. He found that people tend to click on stories that match their views, and platforms respond by recommending similar content, strengthening those beliefs over time.
“I noticed a simple but unsettling fact: two people can open the same news or social media app at the same time and see completely different worlds,” Zhao said.
The study finds that algorithms do not create polarization on their own, but they can amplify it. They build on factors such as political identity, education, geography and social networks, making existing divides more pronounced.
“Social circles, party identity, education levels, economic conditions, geographic sorting and the broader media environment are all among major drivers of polarization," Zhao explained.
The effects can build over time. As platforms learn more about users, recommendations increasingly cater to their political type, raising the likelihood that they will continue to see similar viewpoints.
The study also highlights a difficult trade off for news platforms. Reducing polarization in recommendations may lower user engagement, including clicks and time spent on content. Zhao said platforms may choose to promote more balanced content if they prioritize long term trust or if some users demand a wider range of perspectives.
“When at least one of these two conditions is satisfied (i.e., long term trust or demand for diverse opinion), the platform could adopt a more balanced algorithm design with more objectives,” he said. “It could also treat trust and long-run retention as part of the platform’s success, which helps more diverse content be served.”
In the case of neither condition being met, Zhao suggests government regulation may be necessary, “if the government cares about polarization.”
“From the platform’s perspective, polarization can be viewed as an externality, similar to pollution,” he said. “Regulation could then require platforms to limit the promotion of content that maximizes engagement while creating large negative social consequences.”
The study also identifies who may be most vulnerable. People who already hold strong views are more likely to consume content that aligns with their beliefs, which can further increase their polarization level. Those who rely heavily on a single platform or passively scroll through content may also be more affected.
Zhao suggests that users can take steps to counter these effects by being more intentional in how they consume news.
“Your clicks are not just choices, they are data that facilitate algorithmic personalization,” he notes. “It helps to build a few intentional habits: deliberately follow a wider mix of credible outlets, occasionally read high-quality perspectives you disagree with and use product features that reduce personalization.”
