Singapore AI Safety Red Teaming Challenge

As the use of AI gains traction worldwide, it is increasingly important that AI models navigate regional concerns sensitively, to address local cultures and languages in addition to Western contexts.

 

IMDA organised the inaugural Singapore AI Safety Red Teaming Challenge as an important first step towards making generative AI safer for our region. The Red Teaming Challenge kicked off with an in-person event on 5 November 2024, involving domain experts in culture, languages and AI from 9 Asian countries¹. These experts tested 4 large language models for harmful bias stereotypes about social groups in their countries. This yielded over 1,000 successful prompts in English and regional languages.

 

In the next phase of the Red Teaming Challenge, the online community also got in on the action. From 9-20 December 2024, more than 300 participants (from 7 of the 9 original participating countries) were challenged to generate the same kinds of bias stereotypes.

Stay tuned for updates on the Challenge Findings!

¹The participating countries were: China, India, Indonesia, Japan, Malaysia, Singapore, South Korea, Thailand and Vietnam.