REPURPOSING DEEPFAKES FOR SOCIAL GOOD: THEORY, EMPIRICAL EVIDENCE, AND AI SYSTEMS FOR BIAS MEASUREMENT AND MITIGATION

Abstract

Measuring bias in decision-making represents a pressing challenge facing modern organizations and society. In high-stakes contexts such as healthcare, criminal justice, and employment, even subtle biases can lead to profound and lasting consequences for individuals and communities. To this end, traditional experimental approaches such as audit and correspondence studies have been widely used for bias measurement across various domains. These methods have proven highly effective when applied to textual contexts, where researchers can easily manipulate bias-sensitive attributes like race or age while keeping all other content identical to isolate the causal effects of bias. However, such manipulation is extremely difficult in visual contexts. Prior bias research has relied predominantly on comparisons between natural images of individuals from different demographic groups, which introduces comparability issues. When comparing individuals who differ in race or age, they inevitably also differ in facial structure, background elements, and other characteristics that may legitimately influence decision-making outcomes. These uncontrolled differences make it nearly impossible to determine whether observed differences stem from systematic bias or from other legitimate visual factors. This dissertation comprises four interconnected essays that address these methodological challenges by systematically repurposing deepfakes, a controversial yet technologically promising artificial intelligence technology, for rigorous bias measurement and mitigation. The first essay conducts a comprehensive analysis of 826 academic papers to understand deepfake conceptualizations and develops the theoretical framework of calibration deepfakes for controlled visual bias measurement. The second essay empirically validates this framework through pain assessment experiments with 3,802 assessors, revealing own-race effects and culturally-shaped attentional patterns. The third essay develops Multi-Agent Debiasing Systems, an agentic AI framework that automates the entire bias measurement and correction pipeline. The fourth essay extends the deepfake framework to observational causal inference, developing Deepfake-Informed Control Encoder for Double Machine Learning, and applies it to estimate the causal effect of skin tone on user engagement with 232,089 Instagram posts. Despite deepfakes' widespread association with malicious applications, this research demonstrates how controversial technologies can be transformed into powerful instruments for social benefit when guided by rigorous methodological frameworks.