Publications
The following manuscripts were produced during my doctoral research at ENSAE, Institut Polytechnique de Paris. They apply large-scale computational methods – including NLP, causal inference, and machine learning – to questions of political polarization, media influence, and social mobility. All are currently under review at peer-reviewed journals.
Mediated Cues. How Elite Polarization is Transported Through the Media
with Väinö Yrjänäinen and Måns Magnusson
Do people become more polarized because politicians say polarizing things in the news, or do politicians say polarizing things because people are already polarized? We studied 25 years of French newspaper coverage (n=1,000,000+ articles) and public opinion surveys to find out. The answer: both happen, but it depends on the issue. For some topics, media coverage of elite politicians clearly drives public opinion. For others, politicians seem to follow where public opinion is already going. This means there’s no one-size-fits-all explanation for how political polarization spreads. READ MORE HERE.
How Elite Negativity Shapes Voter Affect: Evidence from the 2021 German Federal Election
solo-authored; currently under review at German Politics
When a political party attacks its opponents on Twitter, does it make voters dislike those opponents more? Using transformer models and OLS regression, I analyzed tweets from candidates in Germany’s 2021 election alongside biweekly voter survey data to find out. The answer is yes: when Party A criticizes Party B on social media, Party A’s supporters report disliking Party B more. The effect is strongest among people who already identify strongly with their party, and it holds up regardless of whether the parties might form a coalition together or how ideologically similar they are. Also, the opposite is not true: a party doesn’t criticize other parties more if they know that their supporters don’t like these parties. READ MORE HERE.
Partisan Tastes Or Ideological Divides? Everyday Interests And Political Identities In A Multi-Party System
with Anastasia Menshikova, Elida Izani Binti Ibrahim, and Miriam Hurtado Bodell; preprint; currently under review at Journal of Computational Social Science
In the U.S., we know that political views often cluster with lifestyle choices – like the stereotype that liberals drive Priuses. But does this happen in countries with more than two major parties? We analyzed what 12,000 politically active Swedish Twitter users follow using mixed membership clustering to see if political affiliation predicts their cultural tastes. It does. Swedish voters’ media consumption, cultural interests, religious views, and even sense of humor align with their partisan identity. Political polarization isn’t just about policy disagreements – it’s seeping into everyday lifestyle choices, even in a multi-party democracy like Sweden. READ MORE HERE.
Media Slant as Political Refraction. Measuring Political Media Slant and Polarization in the French Media Landscape
with Rubing Shen, Arnault Chatelain, and Etienne Ollion
How can we measure whether a newspaper leans left or right without relying on subjective judgments? We developed a new method that analyzes how journalists talk about political issues differently than politicians do. By looking at subtle differences in word choice and framing, we can detect political bias at a very fine level, down to individual paragraphs. We tested this on major French newspapers from 2000-2010 and found that mainstream media became increasingly polarized during this period. This gives researchers and the public a tool to track media bias objectively over time.
Perceptions of Intergenerational Mobility in Germany, Sweden, and the UK: Insights from Machine-Learning Text Analysis
with Alexi Gugushvili and Patrick Präg; preprint; currently under review at European Journal of Sociology
When people compare their lives to their parents’, what are they actually comparing? We used LLMs and clustering-algorithms to analyze thousands of open-ended survey responses from Germany, Sweden, and the UK. While traditional research focuses on income, education, and job status, we found that people think about much more: home ownership, family life, freedom, lifestyle choices, and opportunities. What matters also varies dramatically by country: Swedes emphasize education, Brits focus on housing, Germans talk about freedom and lifestyle. Gender matters too: women are more likely to mention education and family, while men focus on income and career. Understanding what people actually care about when they think about social mobility gives us a richer picture than traditional metrics alone. READ MORE HERE.
Benefit Framing Improves AI Sentiment While Explanatory Schemas Endure
with Alexi Gugushvili
Can brief informational framing change how people think about AI? We ran a preregistered, three-arm survey experiment with a nationally representative sample in Denmark, randomly assigning participants to a benefit-focused frame, a risk-focused frame, or a neutral control. We also used an LLM-executed codebook to analyze open-ended responses on AI’s impact on jobs and on generational dependency. The benefit frame meaningfully shifted overall AI evaluations (+6.5 percentage points more positive, -9.4 points less negative), while the risk frame produced no reliable movement. Neither frame, however, changed beliefs about AI eroding critical thinking skills, and open-ended narratives – predominantly skeptical – were similarly stable across conditions. We conclude that brief informational cues can nudge surface-level attitudes toward AI while leaving the underlying explanatory schemas people use to make sense of it largely intact.