Our project focuses on enhancing the safety and cultural robustness of text-to-image (T2I) models by addressing gaps in existing evaluation datasets. We began by contributing to improving election-related safety at Meta, enhancing AI safeguards against political misinformation and strengthening the accuracy of classifiers. In the process, we identified a lack of research on the composition of open T2I safety datasets and conducted a systematic survey that revealed the concentration of prompts in a few safety categories, along with the underrepresentation of non-English languages and cultural harms. Building on this, in early 2025, we initiated the development of a novel benchmark dataset to evaluate and mitigate biases in T2I models across diverse cultural contexts. Our work aims to provide an open-source framework that enables AI developers to assess and improve the cultural safety of generative models, ultimately enhancing public trust in AI-generated media.
Mentor: Alex Fisher
