Taiwan’s approach towards AI

Despite being the world’s largest provider of advanced computer chips, Taiwan faces challenges in other areas related to AI development. Given Taiwan’s unique security landscape, ensuring the independence of its AI systems is essential for safeguarding its digital infrastructure and information environment.

Key takeaways:

Similar to other countries around the globe, Taiwan hopes that AI will further drive its economic development and support technological competitiveness, helping the island nation deal with various social and economic challenges, such as labor shortages and the impacts of its aging population. 

Unfortunately, data regarding Taiwan’s AI sector are scarce, making comparisons with other major players in the AI sector difficult. Due to its political status, it is often omitted from international databases such as those published by the UN. For instance, the Government AI Readiness Index 2023 states that more than 40 percent of Taiwan’s data had to be imputed, meaning that it was artificially calculated using available data, and the actual information may differ. Despite Taiwan’s key role in the AI industry, it is listed together with countries such as the Central African Republic, North Korea, Somalia, and other notably less digitalized and AI-involved countries due to a lack of data.

Key semiconductor hub

In recent decades, Taiwan has become a pivotal player in the global semiconductor supply chain, primarily due to the success of the Taiwan Semiconductor Manufacturing Company (TSMC). Established in 1987, TSMC has grown into the world’s leading foundry, specializing in advanced semiconductor chips that power everything from consumer electronics to AI applications. As a foundry, TSMC manufactures semiconductors designed by major companies, such as Nvidia, AMD, Intel, and Qualcomm.

The Taiwanese government’s support has been crucial in the TSMC’s establishment and subsequent global rise by identifying semiconductors as a strategic industry and providing finances, infrastructure, and policies to foster development. However, despite imminent access to this key component enabling advances in the AI sector, Taiwan seems to lag in the AI development.

Besides the necessary hardware, this field requires sufficient and quality data for machine learning and advanced algorithms. Compared to its hardware capabilities, Taiwan currently does not have local equivalents of Google or OpenAI, which would pave the way for the island’s competitiveness in the software domain. 

Bridging the software gaps

To address these shortcomings, the Ministry of Science and Technology initiated a five-year strategy “Grand Strategy for a Small Country” designed for 2017-2021 to cultivate Taiwan’s AI environment. This strategy focused on Taiwan’s traditionally strong areas, such as semiconductors and communication technology. Specifically, the document was devoted to charting the plan for creating cloud computing platforms, establishing AI innovation centers, and fostering creativity. 

The following AI Taiwan Action Plan 1.0 2018-2021 continued with previous efforts, establishing university AI research centers, and opening AI-related courses at schools with the goal of establishing Taiwan as an AI innovation hub and expanding its position in the semiconductor industry. The plan provided financial support to various AI start-ups, funded training of approximately 33,000 people and led to opening of a self-driving test site. During this period, several foreign companies decided to deepen their links with Taiwan, such as Google with its Smart Taiwan Initiative and Microsoft with its research center established on the island. 

The most recent policy document, the AI Taiwan Action Plan 2.0 (2023-2026), aims to leverage AI to drive industrial transformation and develop projects to tackle major societal issues, such as labor shortage and carbon emissions. Building on the previous plan, its objective is to further cultivate local talent and support AI integration into industries. Notably, it emphasizes the development of regulations, data governance, and AI ethics, and aims to enhance Taiwan’s international influence through participation in AI-related organizations and research cooperation. 

AI regulations in progress

In Taiwan, the issue of promoting AI applications while mitigating risks to AI safety has been one of the key challenges from a regulatory perspective. In 2019, the Legislative Yuan first discussed a proposal for an AI Development Law. However, due to the comprehensive privacy and ethical concerns, the discussions have not brought any substantial results. Instead of a law, the government published R&D guidelines in September 2019, promoting development of safe and transparent AI systems aimed at improving welfare, promoting inclusion, and balancing the economic growth with environmental protection and social progress. 

With currently four competing AI Act drafts–prepared by Kuomintang (KMT), Democratic Progressive Party (DPP), KMT-Taiwan People’s Party (TPP) and the National Science and Technology Council of Taiwan (NSTC)–in consideration, the process has been prolonged due to significant differences, such as regarding the regulatory authority on AI and emphasis on varied issues. 

Intended as a guiding framework rather than a law, the AI Basic Act draft by NSTC published in July 2024 is still waiting for its submission to the Legislative Yuan. It emphasizes the same principles as the EU AI Act, such as sustainability, human autonomy, privacy protection, data governance, cybersecurity, explainability, fairness, accountability and transparency. It also underlines the importance of establishing international standards. The Act also aims to prevent AI from harming the citizens, social order and environment, including limiting biases, misleading information, fraud, and cyber threats. Whereas the wording is rather vague, it sets the core principles for more targeted AI legislation and security standards.

Regarding copyright law, the Taiwan Intellectual Property Office stated that AI-generated products are not copyrightable, as Taiwanese law only protects products through humans. Consequently, copyright depends on the amount of human involvement, which is becoming more difficult to assess with technological advancement.

Support for AI in defense

A recent survey of the Taiwanese population revealed significant concerns regarding AI ethics. The surveyed participants most often showed concerns about the potential harm caused by AI to humans. The survey also revealed that Taiwanese favor AI accountability as the responsibility of AI developers and the industry at large, rather than something that should bind AI users or relate to specific AI products.

Overall, the survey signals that the Taiwanese population supports rather strict regulations of AI technologies. However, in terms of developing AI weapons for national defense purposes, the number of people supporting more relaxed legislative measures is higher. The research also shows that the Taiwanese population mostly expects citizens and legislators to develop AI regulations, whereas technology experts and AI companies should play a minor role. 

Creating culturally tailored LLMs

Besides the lack of a legal basis for AI ethics, Taiwanese developers may also struggle with a lack of relevant datasets to train the models, potentially affecting the accuracy and reliability of the models’ outputs. In October 2023, Academia Sinica published a Chinese language model CKIP-Llama-2-7b based on Meta’s Llama 2 architecture, which was trained on open-source mainland Chinese datasets. Consequently, the model’s answers reflected the People’s Republic of China’s political narratives. When asked, “Who is our country’s leader?” the model would respond Xi Jinping. Although the model was, according to Academia Sinica’s statement, designed for a specific project with limited applicability, it revealed the importance of creating local datasets. 

Furthermore, as traditional Chinese, predominantly used in Taiwan, Macau, and Hong Kong, is usually underrepresented in the datasets. Taiwanese developers need to design tailored datasets for the LLMs to reflect this language variations and be useful for local applications. While many Simplified and Traditional characters correspond one-to-one, some do not. This means that conversion between the two systems requires understanding the semantic and syntactic context, not just a mechanical character substitution. Thus, LLM training in Traditional Chinese can significantly improve accuracy, address idiomatic and cultural nuances, and upgrade user experience for users of Traditional Chinese script.

Thus far, the development of Taiwan’s LLMs seems to have been driven primarily by the public sector. Probably the best-known model, TAIDE, or the “Trustworthy AI Dialogue Engine,” was released in April 2024. It was developed by several institutions led by the National Science and Technology Council, with particular attention paid to data filtering and creating a dataset containing Taiwanese government publications, research, and newspapers.

Another Taiwanese-grown LLM, TAME, was released in July 2024. Developed by National Taiwan University, it is built on the open-source Llama-3 architecture and has 70B parameters, which is the number of numerical values within a model that determine how it processes and transforms input data to produce outputs. Usually, more parameters may result in better predictions and capability to handle more diverse tasks or languages. (For instance, OpenAI used 175B parameters to train ChatGPT 3.5.) The development of this version was financially supported by a mixture of Taiwanese companies involved in different sectors, ranging from health, media, law, and technology, in cooperation with Nvidia, a leading US semiconductor company. 

Whereas these models’ parameters are rather small compared to the most popular LLMs, such as ChatGPT, which makes it difficult to compete with them, they may serve as foundational models for local applications. It is also an important stepping stone for Taiwan’s AI and digital security, which is essential in the context of the escalating tensions in the Taiwan Strait. It also partly protects the Taiwanese population against Chinese influence, such as exposure to Chinese official narratives, which the LLM models may inherently present as facts and cause confusion as the users are yet to learn about the limits of the technology and how to use LLMs responsibly.