Contributions
- Empirical measurement of ideological homophily in large-scale social networks (10.1 million users, 7 million URLs).
- Quantification of content diversity at each stage of the information diffusion pipeline: network structure, algorithmic exposure, and individual choice.
- Demonstration that user behavior (selective clicking) dominates algorithmic effects in reducing cross-cutting exposure.
- Direct measurement of 903 million News Feed impressions and 59 million clicks on hard news during July 2014–January 2015.
Method
The authors constructed a deidentified dataset of 10.1 million active U.S. Facebook users who self-reported ideology and 226,000 distinct hard-content URLs (news, politics, world affairs) shared by at least 20 ideologically self-identified users. They classified 7 million URLs as hard (news/political) or soft (entertainment/sports) using support-vector machines on unigram/bigram/trigram features.
For each URL, they computed an alignment score (A) as the average ideological affiliation of users who shared it. Alignment captures differences in topic, framing, and slant across user populations, not media slant per se. Conservative-aligned sources (FoxNews.com: A = +0.80) and liberal-aligned sources (HuffingtonPost.com: A = −0.65) showed clear differences.
The measurement pipeline tracked:
- Potential from network: Hard content shared by users' friends (all possibilities).
- Exposed: Content appearing in algorithmically ranked News Feeds.
- Selected: Content users clicked through.
At each stage, they measured the proportion of cross-cutting (ideologically discordant) content.
Results
Friend networks show ideological homophily but substantial bridging. Among liberals and conservatives reporting ideology on Facebook: median ~20% of friendships cross party lines (liberals: 20% conservative friends, IQR [9–36%]; conservatives: 18% liberal friends, IQR [9–30%]). This differs sharply from heavily segregated political blogs but still shows clustering.
Algorithmic ranking has modest effect. If users encountered random content, ~45% of hard news liberals see would be cross-cutting vs. 40% for conservatives. Given friend networks, this drops to 24% (liberals) and 35% (conservatives). The News Feed algorithm reduces this by only 5–8% (risk ratio), showing weak suppression of cross-cutting content.
User choice dominates. After accounting for News Feed position effects, the risk ratio for clicking cross-cutting vs. consistent content is 17% for conservatives and 6% for liberals. This represents a 70% reduction in cross-cutting clicks relative to consistent ones—far larger than algorithmic effects. On average, users clicked on ~7% of hard content available.
Asymmetry between groups. Liberals face fewer cross-cutting opportunities from friends (24% vs. 35% for conservatives) and are less likely to click through even when exposed (6% relative risk ratio vs. 17%).
Notes
Strengths:
- Large-scale, behavioral measurement of actual exposure (not self-report).
- Clear decomposition of network, algorithmic, and choice effects at scale.
- Meticulous methodology for labeling ideological alignment and accounting for confounds (position bias in click rates).
Limitations:
- Limited to active Facebook users who self-report ideology (biased sample: younger, more educated, more female).
- Alignment score reflects distribution of sharers, not intrinsic media slant; topic differences confound with slant.
- Exposure vs. consumption distinction imperfect; users may read News Feed summaries without clicking.
- Findings specific to Facebook's 2014–2015 feed structure; algorithmic ranking has changed significantly since.
Impact and interpretation: The paper's key contribution is showing that individual choice, not algorithm, is the primary bottleneck for cross-cutting exposure—a nuance often lost in later discourse around "filter bubbles." However, the study does not settle whether this is optimal; prior work shows cross-cutting exposure correlates with lower political participation, complicating normative implications.