TLDR: A new AI framework called Latent Topic Synthesis uses large language models (LLMs) and unsupervised clustering to automatically analyze political ads on social media. Applied to Meta ads from the 2024 U.S. Presidential election, it identified key topics, funding patterns, moral framings, and demographic targeting strategies, showing how issues like voting and immigration dominate spending, while abortion and election integrity gain disproportionate reach.
In the ever-evolving landscape of modern politics, social media platforms have become central battlegrounds for shaping public opinion and influencing elections. However, the sheer volume and dynamic nature of content on these platforms pose a significant challenge for researchers and policymakers trying to understand emerging narratives and targeting strategies. A new research paper, titled “Latent Topic Synthesis: Leveraging LLMs for Electoral Ad Analysis,” introduces an innovative framework designed to tackle this very issue.
Authored by Alexander Brady from ETH Zürich and Tunazzina Islam from Purdue University, this paper presents an end-to-end system for automatically generating an interpretable topic taxonomy from unlabeled text corpora. Unlike traditional methods that often require extensive human input or predefined categories, their approach leverages the power of large language models (LLMs) to iteratively build a topic structure from scratch.
The Framework: How AI Deciphers Political Ads
The core of their method, dubbed “Latent Topic Synthesis,” combines unsupervised clustering with prompt-based labeling. It operates in three main stages:
1. Embedding-based Clustering: Political ad texts are first converted into high-dimensional numerical representations (embeddings) using models like Sentence-BERT. These are then reduced in dimensionality using UMAP and grouped into clusters of similar content using HDBSCAN, a hierarchical density-based clustering algorithm.
2. LLM Topic Synthesis: A large language model is then employed to synthesize meaningful labels for these clusters. Starting with an empty list, the LLM iteratively evaluates existing labels and generates new ones as needed, building a comprehensive and semantically rich topic taxonomy.
3. LLM-based Annotation: Once the taxonomy is established, the LLM annotates the cluster representatives with the most appropriate labels. For broader application, a supervised classification model like SetFit can then be trained on these labeled representatives to efficiently categorize the remaining unlabeled ads.
A Deep Dive into 2024 US Presidential Election Ads
To demonstrate the framework’s effectiveness, Brady and Islam applied it to a substantial dataset of over 8,000 unique political advertisements from Meta (Facebook and Instagram), collected just one month before the 2024 U.S. Presidential election. This rich dataset included not only the ad content but also information on funding entities, spending, and impression data broken down by demographics like gender, age, and state.
The analysis uncovered 14 distinct political topics, ranging from “economy” and “voting rights” to “abortion” and “environmental protection.” The framework proved superior to baseline methods like BERTopic in generating coherent and interpretable topic labels, as validated by human annotators.
Also Read:
- New Framework Enhances Detection of Unseen Jailbreak Attacks in Vision-Language Models
- Beyond Binary: New Approach to Align Language Models with Diverse Human Preferences
Key Findings: Spending, Framing, and Targeting
The research yielded several compelling insights into contemporary political advertising strategies:
- Dominant Issues: Ads related to “voting” and “immigration” commanded the highest overall spending and impressions. Interestingly, “abortion” and “election integrity” ads, despite attracting intermediate spending, generated disproportionately high impressions, suggesting strong voter engagement or lower costs per impression for these topics.
- Funding Patterns: Funding for “economy”-related ads was largely dominated by conservative-aligned groups. “Abortion” messaging showed highly polarized spending, with significant investments from both anti-abortion and pro-abortion rights organizations. “Crime/justice” advertising, however, was more decentralized, funded by a mix of PACs and local campaigns.
- Moral Framing: The study also analyzed ads through the lens of Moral Foundation Theory, identifying six core moral dimensions (e.g., care/harm, fairness/cheating, liberty/oppression). Economic appeals often blended multiple moral frames, while “abortion” ads were almost exclusively framed around “liberty/oppression” rhetoric. Strong correlations were found between “fairness/cheating” and “crime/justice,” and “loyalty/betrayal” with “immigration.”
- Demographic Targeting: The analysis revealed precise microtargeting strategies. Younger Floridians, for instance, were more likely to see ads about “affordable housing,” while older Floridians received “abortion”-focused appeals. Similarly, males in Montana were disproportionately targeted with “environmental protection” and “personal freedom” messages, whereas males in Virginia were mobilized around “crime/justice” and “voting.”
This research provides a powerful, scalable, and interpretable tool for understanding the complex dynamics of digital political communication. By shedding light on how political actors segment and appeal to voter demographics, it offers valuable insights for researchers, policymakers, and the public alike. The authors also plan to release the 2024 US presidential election dataset to support future research, which you can find more about in the full paper available at arXiv:2510.15125.
While acknowledging limitations such as reliance on LLM performance and potential biases, the framework represents a significant step towards enabling automated, in-depth analysis of political messaging on social media, helping to monitor emerging narratives and polarization dynamics.
