RAI Overblock Dataset (ROD): A GAN-Based Synthetic Data Generation Approach for Evaluating Overblocking in Responsible AI Systems

Recent deployments of Responsible AI (RAI) systems in high-stakes domains like healthcare and law enforcement have raised concerns about overblocking, where valid inputs are wrongly flagged due to varying levels of RAI checks required for different scenarios. This paper introduces RAI Overblock Dataset (ROD), a novel framework that leverages Generative Adversarial Networks (GANs) to generate synthetic datasets for quantifying overblocking in RAI systems before deployment. We leverage GANs to create realistic data points for evaluating blocking rates of an RAI system before its deployment. Unlike existing methods that rely on real user data and require noise injection or complex privacy-preserving techniques, our approach does not require access to sensitive real user data when evaluating the overblocking issue in RAI systems. This helps better protect user privacy during the RAI system development. By applying fairness metrics across sensitive attributes—such as patient demographics or law enforcement contexts—our approach enables a systematic analysis of RAI decision boundaries to identify and mitigate unintended biases. Our findings highlight how GAN-generated data can be effectively used to improve fairness and accuracy in AI moderation systems, particularly in environments where minimizing bias and ensuring reliability are critical.