Speaker: Louis Wei-Yu Feng, University of Cape Town
Abstract
Existing Large Language Model (LLM) safety benchmarks remain English-centric, severely limiting evaluations for marginalized populations in the Global South. Despite evidence that 85% of women experience online violence, no benchmark systematically evaluates gender-specific harms in African linguistic contexts. A systematic review (PRISMA 2020) of 265 studies (2020–2025) revealed that only 12% tested African languages, and none applied UN Women’s Technology-Facilitated Gender-Based Violence (TFGBV) framework. Findings indicate that safety performance in African languages is 32% worse than in English, with degradation reaching 80%. Current evaluations suffer from data scarcity, Western-centric training, and a lack of gender-aware frameworks. This study argues that existing metrics like Attack Success Rate are inadequate for capturing culturally specific trauma. Consequently, we propose the African Safety Framework (ASFW), which integrates the UN Women TFGBV taxonomy with novel metrics—Cultural Misalignment Rate (CMR), Empathy Score (ES), and Resource Appropriateness Score (RAS)—to advance equitable, context-sensitive AI safety evaluation.
Speaker Bio
Louis Wei-Yu Feng is an interdisciplinary AI researcher and engineer with a background in mechatronics, control systems, and space systems, now specializing in AI safety, fairness, and multilingual large language models (LLMs). His current research focuses on ethical and robust AI for low-resource African languages, emphasizing harm detection, bias mitigation, and adversarial robustness in transformer architectures. With over eight years of combined experience in engineering and project management, Louis integrates systems optimization and community-centered AI development to advance safe, inclusive, and trustworthy technologies.