Nikkei Asia: Taiwan Develops Local Generative AI to Counter Chinese Content Influence

Nikkei Asia reports that Taiwan is accelerating the development of its own generative artificial intelligence (AI) models to counter the influence of Chinese content used in training global AI systems and protect Taiwan's language and culture.

Currently, major global AI models like ChatGPT rely heavily on Chinese-language data sourced primarily from mainland China’s simplified Chinese internet content. In contrast, the amount of traditional Chinese data generated from everyday use in Taiwan is relatively limited.

As a result, when global AI models respond in Chinese to questions about history or society, their answers are sometimes influenced by Chinese political perspectives and censored information. Even in general text generation or search, AI outputs often use vocabulary common in China but less familiar in Taiwan. If this trend continues, the next generation’s understanding of Taiwan’s language and culture may gradually be replaced by Chinese norms.

To address this challenge, Taiwan is advancing the development of the 'Trustworthy AI Dialogue Engine' (TAIDE), a large language model led by Taiwan’s National Science and Technology Council (NSTC).

Huang Han-Hsuan, Associate Research Fellow at the Institute of Information Science, Academia Sinica, stated: 'Without our own AI, we are essentially handing over the entire society’s “brain” to external forces.' He added, 'TAIDE will serve as an AI breakwater, protecting our linguistic, historical, and social autonomy from invisible erosion.'

TAIDE prioritizes local Taiwanese data, including government administrative documents and indigenous languages, rather than collecting vast amounts of Chinese web text and filtering it afterward. The model is built upon high-performance open-source models such as Google’s Gemma.

An NSTC official explained, 'Our strategy is to stand on the shoulders of giants—leveraging overseas technologies to build foundational models requiring massive investment, then combining them with Taiwan’s own data to develop AI with local characteristics.'

The first generation of TAIDE was released in 2024, and the model is expected to integrate NVIDIA’s latest AI chips between August and September 2024, significantly boosting its performance.

TAIDE’s practical applications are expanding, with deployments possible on local servers in government agencies, hospitals, and enterprises. It is already being used to assist in generating official documents such as government memos.

NVIDIA CEO Jensen Huang emphasized that countries should use AI to build their own data and intelligence capabilities, stating that developing AI models reflecting national language, culture, and history is crucial.