A recent landmark study by Singapore’s Infocomm Media Development Authority (IMDA) in collaboration with international AI auditing firm Humane Intelligence has uncovered significant racial, cultural, and gender biases in AI-powered large language models (LLMs). The research, the first of its kind in the Asia-Pacific, tested four LLMs—Meta’s Llama 3, Anthropic’s Claude 3.5 (backed by Amazon Web Services), Cohere’s Aya, and AI Singapore’s Sea-Lion—in English and eight Asian languages including Hindi, Chinese, and Malay.
Out of 5,313 generated responses, more than half were found to be biased. Alarmingly, two-thirds of responses in regional languages exhibited bias, compared to nearly half in English. The LLMs repeatedly reinforced stereotypes: for example, they suggested women were more likely to be scammed online and linked large immigrant groups with higher crime rates. Gender bias was prevalent, with women frequently depicted in caregiving roles and men portrayed as business-oriented. Racial stereotypes emerged as well; one test had the models assign culturally stereotyped names to criminal characters and favor individuals from specific regions when making subjective judgments.
The study involved 54 experts from research institutes across nine countries and 300 online participants who posed everyday, non-provocative questions to the AI systems. The approach aimed to mimic typical user interactions rather than deliberately triggering biased responses. The findings highlight the limitations of current AI guardrails, which were primarily developed in Western contexts, and stress the need for models that better reflect regional cultural sensitivities.
IMDA’s report calls for broader testing and adaptation of AI systems to ensure fairness in applications such as hiring and credit evaluations. It also underscores the importance of international collaboration to improve AI governance and safety. By identifying and addressing these biases, developers can work towards creating more culturally and linguistically sensitive AI tools for a global audience.