Over the last two decades, social media has become a primary news source for millions of people, raising concerns over ‘echo chambers’ and ‘filter bubbles’. This column uses US data covering over a million Facebook articles classified by political slant to show that Democrats and Republicans consume sharply different content on the platform, largely driven by algorithmic curation. The findings highlight how platform design can amplify partisan divides in democratic discourse. Given the finding that extreme articles achieve higher engagement, social media companies have few incentives to curtail the diffusion of more polarised articles.
Over the last two decades, social media has become a primary news source for millions of people. In 2023, for instance, half of US adults reported regularly consuming news on social media platforms like Facebook (Pew Research Center 2023).
The increased reliance on social media for news consumption has generated widespread concern. One of the key worries is that social media’s personalised content and network structure create ‘echo chambers’ and ‘filter bubbles’ that primarily expose users to news that reinforces their existing beliefs (Pariser 2011, Sunstein 2018, Ozdaglar and Acemoglu 2021).
How well do these concerns hold up when tested against data? Measuring polarisation in news consumption is challenging – not least because assessing the slant of individual news articles at scale has proved difficult. In our recent research (Braghieri et al. 2025), we develop a novel method to measure ideological polarisation in news consumption on Facebook, the world’s largest social media platform.
We introduce a content-based approach to classify the ideological slant of individual news articles. We fine-tuned a large language model (GPT-4o) using 3,000 articles rated by experts and then scaled it to more than one million articles published by the 100 most-read US news outlets. This creates a large-scale, content-based dataset of article-level slant. We then used this dataset to quantify polarisation: the difference between the slant of articles engaged with by Democrats versus Republicans on Facebook.
In this column, we present three main findings. First, we explore how article slant varies within and across media outlets. Second, we measure the extent of polarisation in news consumed on Facebook. Third, we identify the mechanisms behind this polarisation.
News production
Before examining how people consume news on social media, it is important to understand the landscape of news production itself. While media outlets vary in their overall ideological leanings, we find that the slant of individual articles within the same outlet can also differ significantly. Figure 1 shows the distribution of article-level slant with a case study of two of the most influential US news outlets: the New York Times and Fox News. While the two outlets clearly differ in their average slant, with the slant distribution of articles from the New York Times skewing visibly more left-wing than that of Fox News, there is a wide distribution of slant within each outlet. In fact, across all 100 outlets we consider in our research, we find that 67% of the overall variation in article-level slant occurs within outlets, and only 33% of the variation occurs across outlets.
Figure 1 Distribution of slant: New York Times and Fox News
Notes: This figure shows the distribution of slant among all hard news articles (N=82,180), published online by the New York Times and Fox News in 2019. The slant scale goes from -2.5 (extremely favourable to the Democratic party) to 2.5 (extremely favourable to the Republican party).
In Figure 2, we show the distribution in political slant across all (approximately one million) news articles published by the 100 most-read US outlets in 2019. Most articles fall near the centre of the spectrum, with 40% classified as neutral. Among those with a clear political lean, left-leaning articles are about twice as common as right-leaning ones. However, right-leaning articles tend to be more extreme on average.
Figure 2 Overall distribution of slant across articles
Notes: This figure shows the distribution of slant among all hard news articles (N=1,065,217) published online by the top 100 US news outlets in 2019. The slant scale goes from -2.5 (extremely favourable to the Democratic party) to 2.5 (extremely favourable to the Republican party).
We next study how slant varies by topic. We classify articles into topics using large language models and calculate the share of left-leaning and right-leaning articles for each topic. Perhaps unsurprisingly, the most over-represented topic among left-leaning articles is the topic of Republican scandals, and the most under-represented topic is that of Democrat scandals. Other topics that are over-represented among right-leaning articles are: technology, the economy, conflicts, and social policy. Conversely, national disasters, the environment, and policing practices are under-represented among right-leaning articles.
Figure 3 Over- versus under-representation of topics among right-leaning articles
Notes: This figure shows which topics are over- or under-represented among right-leaning articles. To construct it, we exclude neutral articles. In the resulting sub-sample of all right-leaning and left-leaning articles, we then compare the share of right-leaning articles in each topic to the overall share of right-leaning articles across all topics. The plotted values indicate how much more or less a topic appears in right-leaning articles relative to the average.
News consumption on social media
To study news consumption on social media, we employ the Social Science One Facebook URLs dataset that provides information about the number of users who viewed, clicked, and shared a given article on Facebook, broken up by users’ political leanings.
We introduce a new measure of polarisation in news consumption, which we define as a normalised version of the difference between the average slant of the articles that Republicans consume on Facebook and the average slant of the articles that Democrats consume on the platform.
Taking our measure to the data, we find that news consumption on Facebook is significantly polarised. The gap between Republicans’ and Democrats’ average news diets is as wide as the average ideological divide between articles from the conservative Washington Times and the moderately liberal Washington Post. As an alternative benchmark, we calculate the slant of articles shared by senators on Twitter/X. We find that the difference in the slant of news consumed by Republicans and Democrats is similar to the difference between articles shared by Republican Senator Marco Rubio and those shared by Democratic Senator Chuck Schumer.
Lastly, using browsing history data from a representative sample of more than 5,000 individuals, we find that the news consumed through social media is substantially more polarised than the online news accessed through other means – for instance, it is more than twice as polarised as news accessed through search engines.
These findings highlight the stark partisan divide in news consumption on social media.
Mechanisms
What drives the high degree of polarisation in news consumption on social media? We investigate four key mechanisms.
First, we study the degree to which the polarisation is driven by ideologically congenial news consumption within outlets. The results are shown in Figure 4. Focusing on the top 20 outlets in terms of online visits, we find that, even within an outlet, the average article consumed by liberal users is virtually always more left leaning than the average article consumed by conservative users. The results in this figure mean that we cannot simply assume that all visitors to an outlet consume the same slant. Instead, it appears as though there are two versions of each outlet, the more liberal version that liberals consume, and the more conservative version that conservatives consume.
Figure 4 Average slant of articles clicked on Facebook by outlet and user political ideology
Notes: This figure shows the click-weighted average slant of articles by Facebook users’ political affinity and outlet for the top 20 US news outlets in 2019. The slant scale goes from -2.5 (extremely favourable to the Democratic party) to 2.5 (extremely favourable to the Republican party).
Second, our measure of polarisation in news consumption on social media can be decomposed into two factors: exposure and selection conditional on exposure. Exposure refers to articles that appear on users’ feeds; selection conditional on exposure refers to the way in which partisans select which articles to click on and read. We find that exposure alone can account for 82% of the degree of polarisation in news consumption on Facebook. In other words, most of the difference in slant between liberals and conservatives already exists when comparing the news they see in their Facebook feed, even before deciding which articles to click. This suggests that social media algorithms, determining which content to show individuals, may contribute to online polarisation.
Third, we ask how the distribution of slant from all articles produced by the top 100 US outlets in 2019 compares to news shared often in the social media ecosystem. We find that extreme articles are approximately four times more likely to be publicly shared on Facebook at least 100 times (see Figure 5). Moreover, opinion pieces, national news, and articles on particularly polarised topics – all characteristics associated with stronger slant – are significantly more likely to be shared as well. The observed patterns of extreme articles achieving higher engagement suggest that, if the objective of social media companies was to maximise engagement, they would likely have few incentives to curtail the diffusion of extreme articles.
Figure 5 Likelihood of being shared on Facebook over 100 times, by article slant
Notes: The grey bars in this figure show the distribution of slant among all hard news articles published online by the top 100 US news outlets in 2019 (left y-axis). The blue line shows the fraction of articles in each bin that were shared at least 100 times on Facebook (right y-axis). The slant scale goes from -2.5 (extremely favourable to the Democratic party) to 2.5 (extremely favourable to the Republican party).
Conclusion
For years, concern about echo chambers on social media has outpaced empirical evidence. This gap stemmed largely from the difficulty of measuring article-level slant. Our work fills that gap by classifying over one million articles and measuring polarisation in what different political groups consume.
We provide strong evidence that the news environment on social media is substantially polarised. This insight is crucial for understanding the dynamics of political polarisation in the digital age and for informing both academic inquiry and policy interventions aimed at fostering a more balanced and informed public discourse.
Source : VOXeu
