Introduction
Microtargeting on social media is a "double-edged sword." While it can facilitate the delivery of relevant health information or career opportunities, it is frequently weaponized to create echo chambers, exacerbate polarization, and nudge individuals toward decisions that may conflict with their own interests.
The primary challenge in addressing this issue is opacity. Social media platforms collect vast amounts of data to construct granular user profiles, yet the public and policymakers lack the tools to understand how messaging is adapted for specific audiences. Often, the sender's true motivation and their perception of the recipient’s demographics remain hidden. This research aims to provide transparency by identifying conflicting messaging choices and the subtle biases introduced when an advertiser speaks to different groups about the same topic.
Methodology
To unmask the "black box" of social media microtargeting, this research employs a multi-pronged computational approach that integrates Natural Language Processing (NLP), graph embeddings, and Large Language Models (LLMs).
Computational Frameworks
- Joint Graph Embeddings: To categorize users even when explicit labels are scarce, the research utilizes models that incorporate both textual information and social interaction data. This leverages social homophily—the tendency for similar individuals to connect.
- Weakly Supervised Learning: Using "model soup" approaches and EM-style iterative processes, the system identifies stances (e.g., pro-energy vs. clean-energy) in political and vaccine-related ads without requiring massive manually labeled datasets.
- Moral Foundations Theory: Messaging is analyzed through the lens of morality, identifying frames such as "liberty/oppression" versus "care/harm" to determine which "moral hook" is deployed for specific demographics.
- LLM-in-the-Loop: By utilizing LLMs in both "human-in-the-loop" and "machine-in-the-loop" configurations, the research discovers nuanced, latent themes and arguments that traditional classifiers might overlook.
Data Sources
The research draws from diverse real-world contexts, including the Facebook Ad Library (2020 U.S. presidential election, COVID-19 vaccines, and climate change), Twitter (X) (lifestyle communities like Yoga and Keto), and Public Discourse Data (immigration and general vaccination debates).
The 'Moral Pivot' and Microtargeting Patterns
A central finding of this research is the "Moral Pivot"—the strategic shift in moral framing based on the target audience.
As illustrated above, when an ad source targets the elderly with vaccine-related content, it often emphasizes liberty and oppression (e.g., framing vaccine passports as government overreach). Conversely, when targeting women of reproductive age, the same source pivots to care and harm (e.g., suggesting vaccines are dangerous for pregnant women).
Other patterns include:
- Geographic Stance Alignment: In climate advertising, "pro-energy" (fossil fuel) ads received the most views in Texas, while "clean-energy" ads dominated in California, mirroring local political and economic identities.
- Latent Theme Discovery: The shift toward LLM-driven analysis allows for the real-time detection of evolving "talking points," moving beyond static categories that quickly become obsolete.
Key Results and Impact
The research successfully developed frameworks to map narratives directly from text to specific arguments (Text → Theme → Argument). Key achievements include:
- Stance Identification: Accurately identified nuanced stances and specific policy issues in the 2020 election cycle using weakly supervised models.
- Holistic Vaccine Discourse Analysis: Integrated three layers of analysis—user stance, underlying reasoning, and responsive morality frames—to provide a comprehensive view of the COVID-19 vaccine debate.
Conclusion and Future Directions
While the landscape of social media is highly dynamic, this research provides a roadmap for identifying how messaging is tailored to demographics. A significant limitation remains the difficulty of quantifying how microtargeting reinforces harmful stereotypes.
Future Work: * Scaling Analysis: Leveraging LLMs to handle the continuous flux of social media messaging. * Equitable Digital Practices: Informing policymakers to ensure digital advertising is fair and transparent. * Human-AI Collaboration: Bridging the gap between technical detection and societal needs through psycho-linguistic collaboration.
Summary updated