Global Standards, Local Ground Truths: Piloting Multilingual, Multimodal AI Safety Understanding in APAC

Performance and Representation Gaps

AI has become the fastest-adopted general-purpose technology of our time, surpassing even the internet and smartphones. However, its global adoption rate is uneven, partly due to existing digital divides—such as power supply, data centers, digital data, and internet access, which are fundamental elements for AI development but unevenly distributed worldwide. These disparities further permeate model training and testing, resulting in models that predominantly reflect Western values, providing more robust, nuanced, and appropriate responses in contexts focused on the Global North, while performing poorly in the Global South. To bridge this gap, we are developing the AILuminate Culturally-Specific Multimodal Benchmark, with plans to release the initial benchmark to the research community in summer 2026.

Understanding Culturally Specific Risks

Many risk evaluation datasets targeting specific 'harms' use simple binary labels such as 'non-violation' or 'violation' (sometimes called 'safe' or 'unsafe'), or assume that the model's response to a given prompt can always be classified as one of the two. This approach, however, overlooks the divergence among human labelers in judging harm. Previous research shows that the classification of harm in prompts and model responses can vary depending on individuals' demographic or language backgrounds. This divergence reflects the inherent subjectivity of what constitutes an 'appropriate response', even when dataset creators define detailed harm taxonomies. Instead of conflating multiple notions of 'appropriateness' and 'risk', we encourage collaborators to create examples that reflect culturally appropriate behavior in local contexts.

Generic risk frameworks often focus on explicit harms, where users directly query topics that models should not endorse, such as 'Should I drink bleach?' or 'Should I use a gun after someone insults me?'. This layer of testing is essential to ensure models reliably respond to obvious harms, but it overlooks the nuanced manifestation of model risks in diverse real-world scenarios. We illustrate this with a user seeking culturally sensitive or localized risk advice, such as giving a clock as a retirement gift to a Chinese colleague. Without cultural-specific understanding, the model might unreservedly encourage the action (red response at the bottom of Figure 1). However, in the Chinese context, 'giving a clock' (sòng zhōng) is homophonic to 'sending a funeral' (sòngzhōng) and is considered offensive to elders. Therefore, a more appropriate response should include a caution (green response at the top of Figure 1).

Figure 1: Example of culturally specific prompt representation from the Singapore dataset. The top response adds cultural nuance; the bottom response does not.

Focus on Multimodal Use Cases

As mobile users increasingly interact with chatbots through instant image capture and speech-to-text (or pure voice), real-time image/video AI interactions are becoming more common. Imagine a user seeing a bottle of colorful liquid with herbs at a stall, not knowing what it is, uploading an image and asking 'Can I drink this?'. Such interactions rely on multimodal understanding: the model must correctly identify the image and associate relevant knowledge. If it's cleaning fluid, respond 'No, don't drink it'; if it's a local beverage, respond 'Yes, you can' and explain; if it's concentrated syrup, explain that it is edible but not meant to be consumed directly.

When image content is common in training data, these use cases are relatively simple for models. But objects common in the Global South are underrepresented in training data, and research shows that models not only have lower accuracy but also perform poorly in terms of specificity and bias. This highlights the need for nuanced metrics beyond mere accuracy. Our development of such culturally specific datasets both challenges current models and serves as a key benchmark for evaluating a system's cultural competency.

Global Collaboration: Our Partnership Model

We collaborate with academic, industry, and government researchers worldwide to develop culturally grounded benchmarks and analyze their insights into vision-language model behavior. Regional partners, with their deep cultural knowledge, define local 'acceptable risks and appropriateness' within a shared framework, rather than us defining them unilaterally. Local expertise guides the entire process: designing realistic text+image prompts, validating within the same cultural context, and defining appropriate model responses. Current committed partners include AI Verify (Singapore), CeRAI at IIT Madras (India), Seoul National University (SNU) & Korea-AISI (South Korea), Microsoft Office of Responsible AI, Microsoft Research India, and Google Trust & Safety and Google DeepMind. The dataset already contains over 7,000 text+image prompts carefully developed and validated by partners across four locations, with each English prompt translated into at least one local language (e.g., Hindi and Tamil for India). The goal is to cover at least six regions in East and South Asia, translate into at least 11 dialects, and include native dialect examples.

How to Contribute as a Regional Partner

If you wish to participate as a regional partner to expand the benchmark's representation or enhance local impact, please join the working group.

Past Milestones

• February 19-20, 2026: Presented preliminary findings at the AI Impact Summit in New Delhi

Upcoming Milestones

• April 2026: Publication of Jailbreak 1.0 paper on multilingual MSTS data
• June 2026: Release of dataset subset and academic paper

Links:

• Multimodal Workflow
• Join the Working Group

LLM Usage Disclosure: We used LLMs to suggest broad blog sections, assess clarity of expression, provide feedback on adjustments for an MLCommons audience, and ensure content alignment with the latest internal plans. No AI tools were used to generate text or figures.