Contributions
- First large-scale, platform-level randomized controlled trial comparing algorithmic vs. chronological timelines for political content amplification
- Systematic analysis of 3,634 legislators across seven countries (US, Japan, UK, France, Spain, Canada, Germany) to measure partisan amplification effects
- Analysis of 6.2 million news articles in the US, using two independent media-bias rating datasets (AllSides and Ad Fontes Media)
- Evidence that algorithmic personalization systematically favors mainstream right-wing political voices across multiple democracies
- Empirical refutation of the "algorithmic radicalizer" hypothesis — no evidence that algorithms preferentially amplify extreme ideologies over moderate ones
Method
The authors leverage Twitter's existing randomized control experiment, established in June 2016 when Twitter introduced personalization. The experiment splits ~58 million accounts (5% of global Twitter users as of June 2020) into:
- Control group (20%): Reverse-chronological feed, no algorithmic personalization
- Treatment group (80%): Algorithmically ranked, personalized feed
The authors define amplification ratio as the proportional difference in reach between treatment and control groups. An amplification of 50% means the treatment group is 50% more likely to encounter a given piece of content. They measure this via "linger impressions" (tweets visible for ≥50% of screen area for ≥500 ms).
For legislator analysis, they identify 3,634 Twitter accounts of currently serving elected officials from public sources (Wikidata, government websites, public Twitter lists). Party affiliation is determined from high-quality public data or the 2019 Chapel Hill Expert Survey. For media analysis, they apply two independent bias rating datasets to classify ~6 million news articles shared on Twitter between April–August 2020.
Results
Legislators (7-country analysis):
- All parties see amplification (0%+ reach gains on personalized timelines), some exceeding 200% (reach 3× larger than chronological)
- Pairwise comparisons of largest mainstream left- vs. right-wing parties in each country show statistically significant right-wing advantage in 6/7 countries
- Largest effects: Canada (Liberals 43% vs. Conservatives 167%), UK (Labour 112% vs. Conservatives 176%)
- At the individual politician level, amplification varies widely within parties (−100% to 400%); permutation tests show no significant correlation between party affiliation and individual amplification
- Far-left and far-right parties (e.g., AfD in Germany, VOX in Spain, La France Insoumise in France) receive lower amplification than moderate/centrist parties in the same country
US media analysis:
- AllSides data: Partisan right marginally more amplified than partisan left; both partisan categories amplified more than neutral sources
- Ad Fontes data: Left sources show lower amplification (10.5%) than other categories; differences among remaining categories less pronounced
- Leave-one-out analysis reveals outliers (e.g., BuzzFeed News, LA Times below category average; Fox News, New York Post above)
Notes
Strengths:
- Massive experimental scale (millions of users, long-running baseline since 2016) and genuine randomized control, not observational
- Multi-country legislative comparison controls for country-specific confounds and political contexts
- Two independent media-bias datasets for robustness
- Clear operational definition of amplification and public-facing metric (linger impression)
- Null finding on extremism amplification is important and counterintuitive — contradicts widespread public belief
Limitations:
- Violates the Stable Unit Treatment Value Assumption from causal inference: control-group users see treatment-group content via retweets, so treatment effects bleed across groups. Authors acknowledge this likely underestimates true causal effects of personalization
- Treatment is not static: personalization algorithm changed over the study period based on performance data, confounding any long-run time trends
- Media-bias classification heavily depends on external datasets (AllSides, Ad Fontes); findings differ materially between them, especially on the left
- Reaches measured in a single country; international audience exposure not analyzed
- Results are aggregate group statistics; individual-level amplification shows enormous variance, limiting generalization to individual politicians
Open questions:
- What mechanism drives the right-wing amplification? Is it content engagement patterns, different posting strategies, interaction styles, or algorithm design choices?
- Does the pattern vary over time? The paper uses one snapshot; does it persist post-2020?
- How do European platform regulations (DSA, NetzDG) and subsequent algorithm changes affect these patterns?
- Why do the two media-bias datasets diverge so sharply on left-leaning sources?