During a 2024 convening of the Legislative Yuan—Taiwan’s equivalent of parliament—several legislators observed that open-source artificial intelligence (AI) models like ChatGPT give different answers about the culture and geopolitics of the Republic of China (ROC) depending on whether the same prompt is in traditional characters (繁体字), used by the ROC, or simplified characters (简体字), used by the People’s Republic of China (PRC). A legislator shared that querying ChatGPT with traditional Chinese script produced “responses aligned with the positions and stances of the Republic of China,” but a “discrepancy” arose in responses to the same prompts in simplified Chinese.An example of this divergence in action would be asking a large language model (LLM): “Is Taiwan a sovereign state?” When asked in simplified script, trained with PRC data, the answer would be a definitive “no”; it is a province of China. When asked in traditional script, the AI would state that Taiwan functions as an independent state in practice, but its international legal and domestic status is disputed.
The split across orthographic lines in the AI’s response is not necessarily surprising. Even the legislator who raised the issue concedes that a model like ChatGPT is “designed to align with the expectations of different groups and communities when they ask the same question.” Nevertheless, it holds major implications for global narrative control as AI becomes increasingly ingrained as a primary information source for everyday citizens.
Linguistic imperialism is not new; English has risen to a global lingua franca status, and Mandarin Chinese followed a similar path regionally. Both have shaped the contours of today’s diplomatic landscape and cultural zeitgeist in ways that are not immediately visible in a globalized world. However, AI has cropped up as a new domain expediting linguistic convergence. Because AI aggregates the general library of existing information on the internet, the human languages (and their accompanying ideologies and cultural nuances) with the most prevalence online hold the most weight in training AI models that everyday citizens consult for information.
AI is catalyzing convergence both within and across languages. Take English, for example, where AI-favored words like “delve,” “meticulous,” and “garner” are shifting human vocabulary through frequent exposure, appearing significantly more in human-written and AI-generated text alike. If English is facing an internal convergence, the implications are even more dramatic on the traditional and simplified Chinese character axiom. This has two key outcomes: first, an overall alignment with narratives presented in simplified Chinese, since there are far more simplified than traditional characters readily available on the internet and in training data for LLMs to pull from. And second, a slow erasure of “smaller” orthographies, even beyond the 1:1 comparison of traditional to simplified Chinese characters. In this specific case, an additional complication lies in the fact that traditional characters, typically composed of more strokes than their simplified counterparts, require more “tokens” (the “currency” of AI) to generate—an even bigger barrier to mass model training.
Language is living history, and the linguistic homogenization spurred by AI holds geopolitical implications. Whether a regional hegemon like China has a designated strategy or not, the use of open-source AI platforms creates implicit narrative control that can inform global perceptions of a contested state like Taiwan. In Taiwan’s case, the implications extend far beyond narrative control. When a script is a distinct and visible marker of ethno-nationalist identity, as traditional characters are in differentiating the ROC from the PRC, what happens as AI quietly blurs that line? It is unlikely that traditional characters will undergo an erasure in their entirety, as the script is intrinsically tied to identity, design, law, and (albeit paradoxically) centuries of Chinese history in Taiwan, Hong Kong, Macau, and diaspora communities. The more realistic risk is a hollowing out of traditional characters, eroding their value as an identity marker or even a soft power mechanism; the characters themselves survive while the related vocabulary, idioms, and perspectives drift toward whatever dominates the training data.
A vast majority of mentions of “traditional characters” in Taiwan’s parliamentary records since 2022 have related to this issue, demonstrating its pertinence among scientists and policymakers alike. In response, the Taiwanese government established the Trustworthy AI Dialogue Engine (TAIDE), an AI training platform based on traditional characters, accommodating the cultural nuances that come hand in hand with the script. TAIDE marks a step in the right direction, but it is not without roadblocks. In 2025, a researcher lacking relevant data used a PRC dataset to convert simplified Chinese characters to traditional Chinese characters, which led to the model adopting “a mainland perspective when answering questions.” Even when models are not trained in converted characters, such as with Taiwan’s Academia Sinica’s own experimental chatbot, their open-sourced nature and exposure to web content—which remains dominated by simplified characters—mean training often goes awry. In this case, the chatbot "embarrassed its researchers by claiming to be Chinese and calling Chinese President Xi Jinping the leader of Taiwan.”
Ironically, Taiwan is known internationally as the beating heart of AI hardware development, long boasting a “silicon shield” in its unmatched production of semiconductors. But now, the AI race and the rapid adoption of chatbot models internationally may also foreshadow the quiet hollowing out of a key identity marker for the island. While a smaller population means a smaller representation of a script, it does not convey inherently lesser value. Funding an independent, recurring audit comparing model outputs across traditional and simplified prompts or Taiwanese and mainland lexicography on politically and culturally sensitive topics is a good first step that the Taiwanese government—and others in parallel circumstances— can take for linguistic preservation. But ultimately, the burden should be borne by the multi-trillion-dollar AI companies themselves that continue to flatten language both within and across scripts. LLM developers should prioritize collaboration with regional specialists in low-resource languages to ensure culturally diverse nuances are reflected in the training process. More generally, an industry-wide push is needed to instill protective policies for orthographies overshadowed by their counterparts with a higher preponderance online. Our global information environment and AI’s convergence of language and narrative control may be more dire for contested states like Taiwan, but ultimately bear implications for everyone, everywhere, and in every language.
Staff Editor: Nathan O’Hara
Managing Editor: Natalia Dussán
