Research
Research interests: media, political economics, culture and identity, AI and information markets
Methods: Causal inference, unstructured data, natural language processing, machine learning
Working Papers & In Progress
with Gabriele Gratton, Pauline Grosjean, and Hasin Yousaf
CEPR Discussion Paper, 2026PDFOnline AppendixInteractive Versionnew
Abstract
Do politicians polarize public debates? We collect 4.75 million tweets debating 57 U.S. mass shootings. We exploit variation in the timing of politicians’ first interventions into each debate and show that partisan and issue-polarized language increases after a politician first tweets about the event. These changes are not preceded by trends in polarization or debate activity, and the timing of political interventions is unrelated to event characteristics or the timing or content of tweets from other celebrities or news organizations. Comparable interventions by non-political celebrities generate no such effects. Politicians’ interventions are themselves unusually partisan, and interventions supplying partisan language produce larger subsequent shifts in both partisan and issue-polarized language. Our findings suggest that politicians supply partisan frames that shape subsequent public discourse and point to elite communication as a supply-side source of polarization on social media.
Press (2)
with Elliott Ash, Sergio Galletta, and Federico Masera
Working Paper, 2026PDF
Abstract
Online creators have become a central source of political information for young adults in recent years, and over the same period, young men have shifted toward the political right relative to young women. We examine whether these two major developments are connected by focusing on “The Joe Rogan Experience”, the leading long-form podcast, which has a disproportionately male audience. Our identification leverages the timing and location of Ultimate Fighting Championship (UFC) events, for which Rogan has long served as the lead commentator. In a difference-in-differences design, we find that local UFC events trigger a significant increase in Google searches and in time spent consuming Joe Rogan episodes among young men. Applying text analysis to podcast transcripts, we document a pronounced shift in Rogan’s political rhetoric: until 2020, Rogan frequently expressed support for Sanders with left-populist rhetoric; starting in 2021, his commentary became supportive of Trump with a right-populist slant. Using administrative voter files, we show that exposure to UFC events until 2020 increased Democratic voting in primaries with Sanders, while after that, exposure to UFC events increased Republican voting in primaries with Trump with all effects concentrated among young men.
Pink Slime News and AI-Mediated Information Markets
with David Zollikofer
Abstract
Investigating networks of commercially motivated pseudo-news websites in Switzerland, using computational methods to trace content production, ownership, and advertising relationships. Exploring how LLMs retrieve and amplify commercially motivated information, with implications for AI reliability, information credibility, and the economics of digital advertising.
The Economics of Visual Journalism
Solo-authored
Socioeconomic Paths into Misogynistic Online Cultures: Evidence from Reddit
with Elliott Ash and Alexander Hoyle
Other Projects
with Matthew Bonci and David Zollikofer
Methodology Report, AI Philosophy Competition, 2026Interactive ReportPDF (more detailed)Video
Abstract
Can AI systems originate genuinely novel research ideas, rather than merely write convincing academic papers? We developed a literature-grounded AI research pipeline that maps academic debates, identifies unanswered questions, and generates candidate theses both within and across research fields. Starting from 32,685 philosophy articles, the system uses citation networks to identify debates, then develops promising ideas into full papers through iterative screening, independent AI peer review, and editorial revision. We evaluate the resulting papers through blind comparisons with published scholarship, including held-out papers inaccessible to the AI authors. Humans designed the research architecture and evaluation framework; AI generated the substantive arguments and wrote the papers.