Taiwan Safety Benchmark and Breeze Guard: Toward Trustworthy AI for Taiwanese Mandarin

Cultural context matters for AI safety: Global safety models perform well on standard benchmarks but have "systematic blind spots" when it comes to region-specific risks like Taiwanese financial scams and culturally embedded hate speech, demonstrating that one-size-fits-all AI safety approaches miss important local harms.

Taiwan-specific harms require dedicated evaluation: The TS-Bench benchmark contains 400 human-curated prompts covering critical domains including financial fraud, medical misinformation, social discrimination, and political manipulation unique to Taiwan, providing the first standardized way to measure whether AI systems can actually handle these localized risks.

Language models need cultural grounding from the start: The research shows that "safety fine-tuning alone is insufficient to introduce new socio-linguistic knowledge from scratch"—meaning AI systems need to be built on foundations that already understand local culture and language, not just patched with safety training afterward.

Specialized models outperform generic ones on local safety: Breeze Guard, a safety model built specifically for Taiwanese Mandarin, significantly outperformed the leading general-purpose 8B safety model (Granite Guardian 3.3) on Taiwan-specific harms, with particularly large improvements in high-context categories like scams.

Responsible AI requires local expertise and investment: Developing trustworthy AI for specific regions demands dedicated resources—human-curated benchmarks, culturally grounded base models, and region-specific safety training—rather than relying on globally trained systems to handle local complexities.