AI've token too much! — Can benchmarks give Taiwan its AI voice back?

Every time a Taiwanese user asks an AI chatbot what to eat for breakfast and receives the answer "hamburger," a small but telling distortion occurs. The AI is not malfunctioning — it is simply reflecting the cultural assumptions embedded in the models that power it.

That observation, offered by Chien Lee-feng (簡立峰), former Managing Director of Google Taiwan and current independent board director at AI startup Appier, cuts to the heart of a structural debate over AI sovereignty. Speaking at a press briefing on Wednesday, March 25, Chien argued that Taiwan must treat AI infrastructure as a matter of national resilience, not merely technological convenience.
（Related： Taiwan Is Running Out of People. A Former Google Taiwan Chief Says It Must Go Global ｜ Latest ）

How AI Models Shape — and Distort — Local Decision-Making

Of the roughly 7,000 languages spoken worldwide, AI systems have been meaningfully optimized for only a handful, Chien noted. More than 6,000 languages remain inadequately served.

The problem, however, goes beyond language coverage. Even when a user queries an AI in Traditional Chinese, the system converts that input into a token sequence — a symbolic representation — before running its computations on a model built primarily around English-language data. The result, analysts argue, is that the answer a Taiwanese user receives may be technically in Chinese but culturally foreign.

Chien illustrated this with a pointed example: ask a mainstream AI chatbot "When is National Day?" and it may return October 1 — the date of the People's Republic of China's National Day — rather than October 10, Taiwan's National Day. The correct query, he argued, is "When is the Republic of China National Day?" Users who do not know to frame questions this way risk receiving systematically skewed answers.

"You think you're asking in Traditional Chinese," Chien said, "but the model's origin determines the answer. You forget there are other options."

The Three Core Distortions: Translation, Quality, and Cultural Bias

Chien identified three structural problems arising from how current AI tokenization works.

The first is translation distortion. When Traditional Chinese text is processed, it is first converted into a token sequence before computation. This intermediate step introduces errors that would not appear in English-language queries.

The second is quality gap. Because most large language models are optimized around English, Chinese-language queries consistently yield lower-quality responses than equivalent English ones — a disparity that is structural, not incidental.

The third, and most consequential, is what Chien termed cultural conditioning. Knowledge is transmitted through token-based representations. Because the dominant AI models are controlled by American and Chinese technology companies, local cultural references, legal norms, and consumer habits are systematically underweighted. The breakfast recommendation is a minor example; in more consequential domains — healthcare guidance, legal information, civic knowledge — the distortions could be far more significant.

Chien also flagged an emerging shift in consumer behavior: an increasing number of users no longer browse platforms like Amazon directly, but instead ask ChatGPT or Gemini to recommend products. This change, he argued, renders conventional brand marketing and user experience design obsolete, forcing companies to rebuild their strategies around AI intermediaries.

What Is a Token — and Why Does It Matter for Sovereignty?

Chien explained that AI systems process language as sequences of tokens — symbolic units that carry contextual meaning through their statistical relationships with other tokens. The Chinese word for "university" (大學) and the English word "University" may look entirely different, but they share similar positional relationships with tokens like "professor," "student," and "campus" in both languages, allowing the model to treat them as conceptually equivalent.

This architecture means that the cultural and linguistic assumptions built into a model's training data are not neutral. They are encoded at the most fundamental level of how the system processes meaning. （Related： Taiwan Is Running Out of People. A Former Google Taiwan Chief Says It Must Go Global ｜ Latest ）

Benchmarks as a Strategic Tool for Smaller Nations

Looking ahead, Chien outlined a scenario in which AI systems manage nuclear power plants, water and electricity infrastructure, transportation networks, and medical decision-making — all without the host country possessing its own foundational model. In such a scenario, he argued, national sovereignty over critical infrastructure becomes structurally compromised.

The solution he proposed does not require Taiwan to build a competing large language model from scratch. Instead, Chien argued, Taiwan must assert the right to define what constitutes a correct answer — through the establishment of localized benchmarks.

Chien compared benchmarks to examination papers. He argued that Taiwan should develop benchmark tests grounded in local law and values — covering areas such as civil and criminal law — and require international AI providers to pass them. Where a model fails, the government should be empowered to demand corrections from vendors.

He pointed to Japan as a working model. When Japan signed a cooperation agreement with OpenAI, Chien noted, the first clause required the company to provide localized benchmarks. He argued Taiwan possesses sufficient economic leverage to make similar demands of international AI vendors.

The goal, as Chien framed it, is ensuring that AI deployed in Taiwan can "speak in Taiwan's terms and abide by Taiwan's laws" — not as a matter of nationalism, but as a prerequisite for institutional resilience in an era of AI-mediated decision-making.

Why This Matters Beyond Taiwan

Taiwan's position reflects a broader tension facing smaller economies and linguistic communities worldwide. As a small number of AI model providers — concentrated in the United States and China — come to mediate an expanding share of information access, consumer behavior, and public services, questions of AI sovereignty are becoming structurally inseparable from questions of digital autonomy and cultural self-determination.

The benchmark framework Chien advocates represents one potential mechanism through which nations without the resources to build foundational models might nonetheless preserve meaningful influence over how AI systems operate within their borders. Whether such mechanisms prove effective at scale remains, for now, an open and contested question.

（Related： Taiwan Is Running Out of People. A Former Google Taiwan Chief Says It Must Go Global ｜ Latest ）