Title: Tackling Online Threats to Democracy with AI in High-Stakes Low-Resource Contexts

Date: June 5, 2025

Time: 9:00 AM - 11:30 AM EST

Location: Hybrid

In-person: CODA C1115 Druid Hills

Virtual: Join Zoom Meeting

https://gatech.zoom.us/j/96147988023?pwd=TXhH2c861qhgrzlLru3t5uRc4cb7V3.1

Daniel K. Nkemelu

Ph.D. Candidate, Human-Centered Computing

School of Interactive Computing

College of Computing

Georgia Institute of Technology

Committee:

Dr. Michael L. Best (co-advisor) – School of Interactive Computing and Sam Nunn School of International Affairs, Georgia Institute of Technology

Dr. Irfan Essa (co-advisor) – School of Interactive Computing, Georgia Institute of Technology

Dr. Ellen Zegura – School of Computer Science, Georgia Institute of Technology

Dr. Munmun De Choudhury – School of Interactive Computing, Georgia Institute of Technology

Dr. Neha Kumar – School of Interactive Computing and Sam Nunn School of International Affairs, Georgia Institute of Technology

Dr. Monojit Choudhury - Mohamed bin Zayed University of Artificial Intelligence

Abstract:

The spread of hate speech, disinformation, and harmful online narratives threatens democracy, especially in low-resource, high-stakes contexts where civil society organizations (CSOs) lack the capacity to respond to these threats effectively. The inability of CSOs to monitor and intervene in these contexts worsens societal vulnerabilities and weakens trust in democratic processes. AI algorithms and tools have the potential to support under-resourced civic actors as they grapple with these challenges. However, their benefits have not been extensively demonstrated in real-world contexts and present a gap in the literature. 

This dissertation demonstrates across multiple studies that AI can support under-resourced civic actors, such as social media monitors and fact-checkers, as they monitor and respond to harmful content online. I draw on the democratic resilience theory as a lens that centers civic society as key stakeholders in democracy building and adopt a participatory AI methodology to confront these challenges in collaboration with stakeholders in Myanmar and Nigeria. I begin by unpacking the manual, time-consuming, and costly nature of hate speech data work in Burmese. I show that despite significant effort, too little usable data was produced to support machine learning modeling. I built on these findings to develop a context-sensitive hate speech data augmentation method. I demonstrated that the novel approach saves significant time and cost and produces reliable data for training hate speech detection models. Next, I evaluated a real-world deployment of a machine learning system to support social media monitoring during the Myanmar 2020 Presidential elections and demonstrate, through a mixed-method study, a significant post-model deployment increase in the proportion of actionable hate posts flagged by monitors.

I conducted an interview study with fact-checkers in Nigeria to investigate existing AI use in fact-checking workflows and identify gaps for AI augmentation. The findings from this study revealed minimal integration of AI tools in fact-checking workflows and surfaced a range of technological, organizational, and environmental barriers to adoption. In response, I developed ClaimFinder, a high-precision AI agent-based system to support fact-checkers in sourcing check-worthy claims from social media. ClaimFinder evaluation results show high accuracy on a social media dataset and strong alignment between the system's rankings and fact-checkers' expert judgment.

This dissertation emphasizes that AI can play a significant role in augmenting civic actors' workflows and helping them respond more effectively to online threats. I argue, however, that AI must be seen as one component within a broader ecosystem of democratic strengthening efforts, not as a stand-alone remedy for these complex social problems.