Red Teaming Challenges

Did you Know Singapore is leading the charge in AI safety by launching the first-ever regional AI Safety Red Teaming Challenge to ensure AI models are more attuned to the diverse cultures and languages of Asia? Even as AI becomes a powerful global force, safety evaluations have often focused on Western contexts, leaving regional harms underrepresented.

This Challenge was a first step towards addressing key gaps and developing a common red teaming methodology. Backed by AI players like AI Singapore, Anthropic, AWS, Cohere, Google, and Meta, the initiative will help make AI safer for the region. 

Participants from nine countries across Asia — experts in fields like sociology, cultural studies, and linguistics — red teamed four large language models (LLMs) to spot bias stereotypes directed at social groups in their countries, both in English and regional languages. By involving these experts, the Challenge aims to deepen our understanding of the diversity of AI harms and how well AI models perform in different contexts. 

This initiative to create safer and more culturally sensitive AI models will wrap up with the publication of an evaluation report in early 2025, to drive further AI safety research.