On (Mis)Information

Author: Linegar, Mitchell

Year: 2026

Degree: Dissertation (Ph.D.)

Advisor: Alvarez, R. Michael

Committee Members: Katz, Jonathan N.; Mazumdar, Eric V.; Martin, Gregory; Alvarez, R. Michael

Option: Social Science

Abstract

This dissertation studies how American voters relate to political information, in three threads. The first asks who believes misinformation about elections. Three representative national samples fielded around the 2024 U.S. presidential election show belief in specific fraud claims to be widespread and partisan-asymmetric; a 16-year SPAE pool shows the asymmetry is structural rather than transient; and a standardized misinformation susceptibility test reveals that susceptibility concentrates along partisan, educational, and trust lines.

The second thread pushes against the harms of that misinformation. AI-authored prebunking articles reduce belief in election rumors at machine speed without per-rumor human cost, and chatbot conversations carrying the same content lift institutional confidence in the audiences whose trust is most fragile. A two-domain experiment with 4{,}338 voters introduces \emph{Pre-Bolstering}: the proactive defense of institutional trust before persuasive rumors target the institution.

The third thread extracts political information from long texts at scale. C-TreePO treats long-document coding as an oracle-preserving compression problem on a tree, with the alignment between LLM labels and the analyst's own intuitions auditable at the node level. On the Manifesto/Benoit benchmark of party-platform policy positions, C-TreePO clears matched open-weight and proprietary LLM baselines. Misinformation labeling is one natural application; the framework runs well beyond it.

Together the three threads describe how American voters and democratic institutions relate to political information: who believes its sharpest false forms, what to do about that belief, and how to measure the long texts through which information of all kinds travels.