Using AI to Summarize US Presidential Campaign TV Advertisement Videos, 1952-2012

Adam Breuer, Bryce J. Dietrich, Michael H. Crespin, Matthew Butler, J. A. Pyrse, Kosuke ImaiarXiv 2025 · DOI · arXiv

Introduction: The Challenge of Historical Ad Analysis

Political scientists have long recognized television advertisements as a primary medium through which candidates communicate their platforms and personas to the electorate. However, large-scale longitudinal research has been hindered by two significant obstacles: data scarcity and the labor-intensive nature of manual analysis. While digital archives of recent election cycles are readily available, advertisements from the pre-1996 era often reside in physical archives, making them difficult to access and search. Furthermore, the traditional method of "human-coding"—where researchers hire assistants to manually watch and summarize thousands of videos—is prohibitively slow, expensive, and difficult to scale.

To address these challenges, researchers have developed an automated, multi-stage AI pipeline designed to unlock historical archives. This system transforms nearly 10,000 historical videos into machine-searchable transcripts and high-quality summaries, providing a scalable solution for political communication research.

The AI Analysis Pipeline

The core of this research is a four-step, parallelized AI analysis pipeline that processes raw video files into structured data. The system begins with automated preprocessing and trimming using the Google Video Intelligence API to remove non-advertisement content, such as archive metadata or TV station IDs. This ensures that the subsequent analysis focuses solely on the campaign message.

The four-step automated pipeline for processing historical political advertisements.

Following preprocessing, the pipeline executes three primary analytical tasks:

  1. Transcription: Audio is processed using Whisper-large-v3, which generates text and segments the transcript based on natural pauses in speech.
  2. Key Frame Extraction: To capture visual context, the system extracts frames at regular three-second intervals and at the midpoint of every speech segment. This "middle-of-speech" heuristic ensures that the AI captures the most visually representative moments of a narrator's message.
  3. Multimodal Summarization: The pipeline utilizes GPT-4-Vision to generate brief descriptions of each key frame. Finally, GPT-4 synthesizes the full transcript and these visual descriptions into a concise, 50-word academic summary of the advertisement.

Dataset Overview

The primary contribution of this work is the release of the Presidential Ad Dataset, the largest digital collection of its kind. Spanning six decades of American political history, the dataset includes over 100 hours of video content, complete with transcripts and AI-generated summaries.

Summary statistics of the newly released Presidential Ad Dataset.

As shown in the table above, the dataset comprises 9,707 ads from the 1952 Eisenhower campaign through the 2012 Obama-Romney election. The collection is heavily weighted toward Democratic (5,859) and Republican (3,848) advertisements, providing a robust foundation for comparative partisan analysis across different eras of media technology and political rhetoric.

Performance and Validation

The researchers validated the pipeline's output through both technical metrics and human expert evaluation. The transcription process achieved a Word Error Rate (WER) of 0.0329, indicating a level of accuracy comparable to professional human transcription.

In a blind evaluation conducted by political science graduate students, the AI-generated summaries were compared against those written by human research assistants. Using a Mixed-Effects Ordinal Logit (MEOL) model, the study found that AI summaries were statistically superior to human ones. Specifically, the AI summaries had 39.8% greater odds of receiving higher scores for coherence, 85.9% greater odds for fluency, and 39.4% greater odds for relevance. Crucially, this performance remained consistent across different historical eras, demonstrating that the AI is not biased toward modern, high-definition video quality.

Conclusion and Future Directions

By leveraging multimodal AI, this research provides a blueprint for the large-scale analysis of historical video archives. While the current dataset concludes at 2012 due to the increasing difficulty of obtaining ads from modern consultants, the methodology itself is highly portable. The authors suggest that this pipeline could be applied to other massive video repositories, such as news broadcasts or congressional proceedings, enabling a new era of visual and textual analysis in the social sciences. This work not only preserves political history but makes it searchable and quantifiable for future generations of scholars.

Summary updated

Abstract

This paper introduces the largest and most comprehensive dataset of US presidential campaign television advertisements, available in digital format. The dataset also includes machine-searchable transcripts and high-quality summaries designed to facilitate a variety of academic research. To date, there has been great interest in collecting and analyzing US presidential campaign advertisements, but the need for manual procurement and annotation led many to rely on smaller subsets. We design a large-scale parallelized, AI-based analysis pipeline that automates the laborious process of preparing, transcribing, and summarizing videos. We then apply this methodology to the 9,707 presidential ads from the Julian P. Kanter Political Commercial Archive. We conduct extensive human evaluations to show that these transcripts and summaries match the quality of manually generated alternatives. We illustrate the value of this data by including an application that tracks the genesis and evolution of current focal issue areas over seven decades of presidential elections. Our analysis pipeline and codebase also show how to use LLM-based tools to obtain high-quality summaries for other video datasets.
BibTeX
@misc{breuer2025using,
  author = {Adam Breuer and Bryce J. Dietrich and Michael H. Crespin and Matthew Butler and J. A. Pyrse and Kosuke Imai},
  title = {{Using AI to Summarize US Presidential Campaign TV Advertisement Videos, 1952-2012}},
  year = {2025},
  eprint = {2503.22589},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.22589},
}
Figure