Large Language Models as a Detector for Political Deepfakes

Author

• Gracie L. Roberts, University of MississippiFollow

Date of Award

Spring 5-9-2026

Document Type

Undergraduate Thesis

Department

Management Information Systems

First Advisor

Da Xu

Second Advisor

Sumali Conlon

Third Advisor

Xin Dang

Relational Format

Dissertation/Thesis

Abstract

The advancement of generative artificial intelligence has allowed for the creation of highly realistic political deepfakes, posing significant threats to democratic integrity and trust in governmental institutions. This study investigates the effectiveness of Large Language Models (LLMs) as detectors for political deepfakes, specifically exploring their ability to address the transferability, interpretability, and robustness limitations inherent in traditional detection systems.

The research was conducted in four steps including data collection, model selection, experimental testing, and evaluation. Two datasets were used in this study. The first dataset uses 833 fake images from the Political Deepfake Incidents Database (PDID) and 833 real images of politicians sourced from Wikimedia. The second dataset uses 83 real images from Wikimedia and 80 fake images generated by Grok. The images were classified using two models: an open-source Convolutional Neural Network (CNN)-Based detector, EfficientNet-B0, and ChatGPT’s GPT-5 model. The methodology included zero-shot prompting and a subsequent optimization phase that integrated agentic analysis of prior misclassified data and a shift toward semantic reasoning.

Findings indicate that the LLM significantly outperformed the traditional detector, which proved inaccurate on real-world datasets due to its reliance on pixel-level artifacts. The zero-shot method achieved high accuracy on the PDID dataset but relatively lower performance on the higher-quality Grok dataset. Following prompt optimization to emphasize semantic and contextual logic, the LLM’s accuracy on the Grok dataset increased. Interpretability analysis revealed that while the model initially focused on visual cues like lighting and texture, optimization allowed it to identify deepfakes by recognizing historically or procedurally implausible scenarios.

The study concludes that multimodal LLMs offer a distinct advantage over traditional methods by complementing visual detection with semantic reasoning. These models can identify sophisticated synthetic media that may appear visually flawless but remain logically inconsistent within a political context.

Recommended Citation

Roberts, Gracie L., "Large Language Models as a Detector for Political Deepfakes" (2026). Honors Theses. 3499.
https://egrove.olemiss.edu/hon_thesis/3499