My dissertation examines how AI-driven information systems shape political information environments and democratic processes. Across three interrelated projects, I investigate how interventions in online information ecosystems and AI-mediated communication can influence political behavior, polarization, and democratic trust.
Chapter 1: Muting Durably Reduces Engagement with Untrustworthy Online Sources; Media Literacy Doesn’t
Co-authors: Brendan Nyhan (Dartmouth College), Ro’ee Levy (Tel Aviv University), Filippo Menczer (Indiana University), Maria Elizabeth (Betsi) Grabe (Boston University), and Giovanni Luca Ciampaglia (University of Maryland)
Research Question: Can reducing exposure to untrustworthy information sources change the beliefs and behavior of people who actively seek out such content?
Online misinformation remains a persistent challenge despite extensive efforts focused on media literacy and fact-checking. This project asks whether changing the structure of people’s information environments can be more effective than attempting to improve individuals’ ability to evaluate information.
Using a digital field experiment on X (formerly Twitter), we randomly reduced participants’ exposure to untrustworthy sources by muting a subset of accounts they engaged with. We compared this intervention to a media literacy treatment designed to improve participants’ ability to identify misleading information.
Our findings indicate that the muting intervention produces substantial and lasting reductions in engagement with targeted untrustworthy sources. Importantly, we find no evidence that participants compensate by shifting their attention to other unmuted untrustworthy sources. These effects persist even when participants have the option to reverse the intervention by manually unmuting the accounts, suggesting that a one-time, opt-in change to the information environment can generate durable behavioral changes without continued enforcement or incentives.
By contrast, the media literacy intervention temporarily improved discernment but did not meaningfully alter participants’ online engagement behaviors. Together, these findings suggest that structural interventions that reshape information exposure may be more effective at changing behavior than interventions focused solely on cognitive skills.
Chapter 2: Challenging Partisan Expectations Reduces Political Polarization
Co-authors: Ozgur Can Seckin (Indiana University Bloomington), Saumya Bhadani (University of Maryland), Alessandro Flammini (Indiana University Bloomington), Giovanni Luca Ciampaglia (University of Maryland), and Bao Tran Truong (Indiana University Bloomington; Dresden University of Technology)
Research Question: Can exposure to unexpected agreement with outgroup partisans or disagreement with ingroup partisans reduce political polarization?
Political polarization is often sustained not only by substantive disagreements but also by inaccurate perceptions about political groups. Many people believe that members of their own political group largely agree with one another, while members of the opposing political group hold uniformly extreme views.
To examine whether AI chatbots can help reduce these misperceptions, we designed a survey experiment in which participants engaged in brief political conversations with large language model (LLM)-powered chatbots. Using a 2×2 experimental design, we varied both the chatbot’s partisan identity (ingroup versus outgroup) and conversational stance (agreement versus disagreement).
Our findings show that disagreement from an ingroup chatbot and agreement from an outgroup chatbot both reduce political polarization. These effects do not arise because participants change their own policy preferences or become persuaded by the chatbot’s arguments. Instead, participants update their perceptions of political groups. They come to view their own political group as more internally diverse and the opposing political group as less uniformly extreme than they had previously assumed.
These results suggest that AI chatbots may serve as a novel tool for reducing political polarization by reshaping perceptions of partisan boundaries rather than changing substantive political beliefs.
Chapter 3: Can AI Strengthen Democracy? Improving Public Trust in Elections with a Civic Chatbot
Co-authors: Giovanni Luca Ciampaglia, Cody Buntain, Riley Lankes (University of Maryland), and Patrick Wu (American University)
Research Question: Can a civic-oriented AI chatbot grounded in official election information improve voter knowledge, trust, and civic engagement?
Free and fair elections depend not only on institutional integrity but also on citizens’ ability to access reliable information and trust the processes through which elections are administered. Yet many voters face challenges navigating election procedures, identifying trustworthy information sources, and understanding how elections operate—particularly at the state and local levels.
This project examines whether a civic-oriented AI chatbot can improve voter knowledge and strengthen trust in democratic institutions. The chatbot will be grounded exclusively in official election information from local election authorities, allowing users to ask questions about voting procedures, election administration, ballot access, and related topics.
To evaluate its impact, I plan to conduct a three-arm randomized controlled trial during the November 2026 U.S. midterm elections with residents of Montgomery County, Maryland. Participants will be randomly assigned to conditions that vary access to election information resources, including the AI chatbot.
The study will examine whether interacting with the chatbot improves understanding of election procedures, reduces false beliefs about elections, increases trust in election officials and democratic institutions, and promotes civic engagement. More broadly, the project explores how AI systems can be designed to support democratic participation by helping citizens navigate complex civic information environments.