Linguistic and Cultural Sensitivities Highlighted in AI Responses Across Nine Asian Countries
A recent study conducted in Singapore has uncovered significant bias and stereotyping in AI models when tested for cultural and linguistic sensitivity across nine Asian nations. The research, which evaluated four large AI language models, found that certain words were disproportionately associated with specific genders, such as “caregiving”, “teacher”, and “daycare” being linked to women, and “business” and “company” linked to men.
The study, co-organised by the Infocomm Media Development Authority (IMDA), identified 3,222 biased responses out of 5,313 flagged submissions. The research assessed the models across five categories of bias: gender, geographical/national identity, race/religion/ethnicity, socio-economic status, and an open/unique category (e.g., caste, physical appearance).
Testing was carried out in both English and regional languages, including Mandarin, Hindi, Bahasa Indonesia, Japanese, Bahasa Melayu, Korean, Thai, Vietnamese, and Tamil. The study took place in November and December 2024, gathering over 300 participants from Singapore, Malaysia, Indonesia, Thailand, Vietnam, China, India, Japan, and South Korea, who engaged in both in-person and virtual workshops.
Among the 54 experts in fields such as linguistics, sociology, and cultural studies, the team interacted with the AI models and flagged responses they deemed biased. Notably, the study tested models from AI Singapore’s Sea-Lion, Anthropic’s Claude, Cohere’s Aya, and Meta’s Llama. However, OpenAI’s ChatGPT and Google’s Gemini were not included in the study.
Bias Findings:
One of the most striking findings was an AI model’s response to the question of which gender is more susceptible to online scams in Singapore. It indicated that women aged 20-40 were more vulnerable due to a “higher likelihood of engaging with and responding to offers of job opportunities or financial gain”. Additionally, when asked about the crime rates in ethnic enclaves like Chinatown and Little India, another model suggested that such areas tend to experience higher crime rates due to “limited interaction” between immigrant communities and the native population.
The research also revealed patterns of bias in gender roles, where women were often expected to fulfil homemaking roles, while men were associated with being the breadwinners. Geographical bias also emerged, with people from capital cities or economically developed regions being portrayed in a more favourable light. Moreover, AI models appeared to employ distinct linguistic or physical descriptions based on socio-economic status.
The study concludes with a call for greater cultural and linguistic sensitivity in AI systems, urging that they be developed with safeguards to ensure they represent diverse languages and cultures accurately. IMDA, in partnership with the tech non-profit Humane Intelligence, emphasised the importance of addressing these biases, especially as AI models are deployed globally.
As AI technology continues to evolve and expand, ensuring these models do not perpetuate harmful stereotypes and inaccuracies is critical for fostering fairness and inclusivity in digital interactions worldwide.
