Developing Agentic AI Systems for Social Science Research: Use Case on Generative AI and Disinformation

• In Event: Discrimination, Fairness, and Algorithmic Bias in Machine Learning Applications

Abstract

Overview: This paper introduces an agentic AI system—a coordinated network of specialized AI agents—designed to support the full social science research workflow: literature review, data collection, and mixed-methods analysis. The system embeds human oversight at critical decision points, amplifying researcher judgment rather than replacing it, and freeing scholars from labor-intensive tasks to focus on higher-order analytical work.

The Problem: The motivation for such a system is straightforward. Social science research is increasingly data-intensive. Scholars studying misinformation, policy diffusion, or public opinion must navigate vast, heterogeneous data environments—monitoring dozens of sources, processing non-English content, coding thousands of observations, and synthesizing large bodies of literature. Yet few integrated, researcher-controlled framework currently exist to coordinate these tasks across the research lifecycle.

System Architecture: To address this gap, the system comprises discrete but interoperable agents, each handling a specific research function:
Data Source Identification — Scans fact-checkers, government databases, and civil society repositories to maintain a curated source registry
Automated Data Collection — Retrieves disinformation claims weekly by country and platform (10,000+ claims collected to date)
Full-Text Extraction — Structures claim text, annotations, and metadata for downstream analysis
Human Coding & Model Training — Researchers manually code a representative sample; this labeled data trains an AI agent to extend coding to the full corpus
Mixed-Methods Analysis — Produces quantitative summaries and flags cases for qualitative scrutiny; researchers retain final interpretive authority
Literature Review — Synthesizes scholarship and identifies theoretical gaps for researcher evaluation
Writing — Human authors retain full control; the system supports evidence retrieval and citation management only

Applied Case Study: South Asia

To validate the framework empirically, we deploy it in a substantive research context. South Asia is an understudied but critical case—its multilingual information environment, expanding social media penetration, and histories of ethnoreligious and electoral conflict create ideal conditions for AI-assisted disinformation. We are building the first large-scale, cross-national dataset of AI-assisted disinformation claims from fact-checkers across India, Pakistan, Bangladesh, and Sri Lanka (2022–present). Preliminary data reveals meaningful variation in claim types, platform distribution, and generative AI tool usage across countries and election cycles.

Broader Implications

The insights emerging from this case point to wider applicability. This framework can be adapted to any policy research domain characterized by high data volume, multilingual sources, or mixed-methods needs—including climate policy, public health communication, and regulatory compliance.Ultimately, the goal is not to automate scholarship, but to restructure its labor. As AI capabilities advance, developing methodological norms for their responsible integration into social science is a matter of scholarly integrity—not merely efficiency.

Author

• Shahidur Rashid Talukdar, Virginia Polytechnic Institute & State University
  Presenting Author