Agustin Prinetti

PhD Candidate in Political Science, Rice University. Visiting fellow, Centre for Political Research (CEVIPOF), Sciences Po. Editorial assistant, International Organization.

I am a PhD candidate in Political Science at Rice University. This semester, I will be a visiting fellow at the Centre for Political Research at Sciences Po. I also serve as the editorial assistant for International Organization.

I work on the intersection of comparative politics, communication, and technology. My research explores the adoption of new digital tools, with a focus on AI chatbots, and their influence on the ways citizens learn about and engage with politics.

Originally from Argentina, I began my journey in Political Science at Universidad Torcuato Di Tella. Before starting my PhD, I worked as a Research Analyst for the Government of the City of Buenos Aires.

Outside of academia, I enjoy learning languages: I am fluent in English, Spanish, and Italian, and I currently study French. I also enjoy swimming and rewatching The Simpsons.

Download CV (PDF) · Email · X / Twitter · Bluesky · LinkedIn

Research Agenda

I study how citizens in different countries interact with political content on social media that is deceptive, hostile, or extremist. My research examines how users engage with that content and how it influences their political behavior. Methodologically, I combine the observational study of social media data with immersive online experiments that closely recreate real-world digital environments.

  • Misinformation & Factual Verification
  • Political Manipulation of AI Chatbots
  • Moderation of Extremist Content
  • Campaign Negativity & Attack Advertising

Dissertation

AI chatbots are becoming intermediaries of political information: built to answer any question, they now routinely fact-check, summarize, and explain politics to millions of citizens around the world.

But unlike social media feeds, chatbots are consulted one user at a time, answer in a single authoritative voice, and draw on sources users often do not see.

My dissertation explores how reliable that voice is, how easily it can be manipulated by interested actors, and how it decides which users should hear about politics at all.

Under Review

Bystanders on Social Media: Audience Moderation and the Spread of Political Extremism

Joyce, Conner B., Maranda Joyce, and Agustin Prinetti. Revise and Resubmit at the Journal of Experimental Political Science. Manuscript

What explains participation in crowd moderation on social media platforms? Social media platforms rely on crowd moderation (the reporting of extremist content by users) to identify and remove harmful content. We argue that participation in crowd moderation can be explained by the Volunteer's Dilemma Game, where users' decision to report is a function of the benefits, costs, and number of other observers. We expect that when users encounter a greater number of extremist posts or they believe that others will report extremist content, their inclination to participate in crowd moderation is depressed. We test these expectations by employing a novel experimental approach in which subjects interact with social media feeds that simulate online environments. Our study would help explain crowd moderation's shortcomings and the widespread circulation of extremist content on social media platforms.

Designing for Power: Covariates, Indices, and Efficiency in Survey Experiments

Peterson, Erik, Matthew Tyler, and Agustin Prinetti. Revise and Resubmit at Political Science Research and Methods. Preprint (SSRN)

Survey experiments are widely used across the social sciences, yet many are underpowered. We examine two well-known, but unevenly applied, strategies for improving statistical efficiency: covariate adjustment and multi-item index outcomes. We formally evaluate when these practices increase power and revisit over four hundred recent social science survey experiments to empirically assess their prevalence and effectiveness. We show these techniques reduce distinct sources of variance and function as complements, not substitutes, for increasing power. However, the majority of survey experiments omit these practices or employ them ineffectively by adjusting for generic demographic covariates that yield minimal precision gains. Only a small minority of studies use specialized covariates or index outcomes. Fewer than five percent combine both practices. Drawing on simulations based on recent studies, we argue designs combining specialized covariates with a short index outcome offer a practical and robust way to increase the statistical power of survey experiments.

Do Androids Dream of Corrective Tweets? The Reliance of AI Factual Verification on Professional Fact-Checking

Prinetti, Agustin. Preprint (SSRN)

Can AI chatbots replace humans in the fight against online political misinformation? As partnerships between professional fact-checkers and social media companies erode, platforms have started promoting AI-based replacements. Motivated by political communication research documenting a trade-off between speed and accuracy in fact-checking, I argue that AI tools may exacerbate this pattern. I focus on Grok, the AI chatbot integrated into X (formerly Twitter) and promoted as a factual verification tool. Through an observational study of Grok's verifications over time and a controlled design that manipulates its access to external sources, I show that Grok's accuracy increases as professional fact-checking articles become available, and that its access to them is a substantial driver of its performance. These findings suggest that AI should be understood as a complement rather than a substitute for fact-checkers, as its accuracy depends, to a relevant extent, on the same human activity it seeks to replace.

Working Papers

Artificial Infiltration: How Low-Quality Sources Shape AI Chatbots' Political Advice

Prinetti, Agustin. In progress.

Citizens increasingly rely on AI chatbots to learn about politics, yet these tools read the same web that malicious actors populate with fake news outlets. Can low-quality sources infiltrate the political advice chatbots give? I answer this question with a preregistered field experiment that replicates the logic of an influence operation under a sanitized protocol. I created three websites posing as independent outlets and published short articles that framed verifiable facts about 5,468 elected representatives across levels of government in three countries in either a positive or negative direction. As I sequentially published and removed the articles, I repeatedly queried Grok, ChatGPT, and Gemini to evaluate each politician. Only Grok retrieved the planted articles, doing so far more often at more local levels, and its assessments shifted with the valence of the coverage. In a second wave, retrieval disappeared, but the planted pages were cited again when Grok's access to certain media sources was restricted. Moreover, using georeferenced data on local media outlets in France, I show that infiltration was more likely in districts with fewer local outlets, even after controlling for district size. The findings suggest that while this manipulation of AI political advice is not universal, it is still possible at very low cost. It also identifies a new consequence of the decline of local media, as news deserts may leave AI chatbots more exposed to manipulative sources.

Research Funding

Institutions that supported me with grants and fellowships to conduct and present my research during my PhD: Rice's School of Social Sciences; the Fulbright Foreign Student Program, US Department of State; the Chateaubriand Fellowship, Embassy of France in the US; the American Political Science Association; the Southern Political Science Association; and Rice's Center for Computational Insights on Inequality and Society.

Teaching

I have worked as a TA for graduate-level courses on statistics and quantitative methods, and for undergraduate-level substantive classes, both in English and Spanish. I have experience leading discussion sections and evaluating students' performance through assignments and exams. In 2024, I completed Rice's Graduate Certificate in Teaching and Learning.

  • Introduction to Maximum Likelihood Estimation I and II (PhD level, Rice University, Fall 2025 and Spring 2026, Prof. Randy Stevenson; average student rating 5.0/5)
  • Quantitative Applications in Global Affairs (Master's level, Baker Institute / Rice University, Fall 2024 and Fall 2025, Prof. Daniel Potter; average student rating 4.9/5)
  • Introduction to American Politics (Undergraduate level, Rice University, Fall 2023, Prof. Jaclyn Kaslovsky; average student rating 4.4/5)
  • Comparative Political Systems (Undergraduate level, Universidad Torcuato Di Tella, 2021, Prof. Ana Maria Mustapic; average student rating 4.3/5)