Japan’s AI Sovereignty Experiment as Seen through Shisa.ai
A while ago, I spoke with a foreign friend who has been a longtime entrepreneur in Japan, and we talked about his local experience. He said that what takes the most time to get used to is Japan’s unique way of communicating, as opposed to competing in the market.
“Sometimes they talk very politely, but I’m still not sure if it’s really an offer or a refusal.” He gave a wry smile and recalled a time when he asked a client for help in introducing a contact, and the other party politely replied: “考えさせてください” (Please let me think about it). He thought he would be contacted again, but that was the end of the letter.
Such a tone of voice is actually a common Japanese euphemism for refusal—”It’s not that I won’t help, it’s just that it’s not convenient to say no.” For those unfamiliar with Japanese culture, this style of communication may easily lead to misunderstandings.
This just resonates with me so much. Because language is never just about translating words, but about tone, context, and mutual understanding.
I can’t help but think of the AI Large Language Models (LLM) that we now use every day. Even entrepreneurs who have lived in the country for a long time can misunderstand each other’s meaning, so how can AI models trained in the “English context” really grasp the nuances of these cultures?
AI’s Cultural Bias
In fact, this question has recently been confirmed by a study from MIT Sloan.
The team experimented with large-scale language models, such as OpenAI’s ChatGPT and Baidu’s Wenxin Yiyan, and found that the models showed different cultural biases when answering in different languages. When prompted in Chinese, the AI tends to show a “group-oriented” mindset, whereas in English, the AI tends to express itself in an “individual-oriented” way.
For example, when prompted to design an insurance advertisement slogan and the input is in Chinese, the AI model may produce something like: “Your family’s future is your promise.” In contrast, the English input might generate “Your future, your peace, our insurance.” The same question, in different languages, reflects a very different ordering of cultural values.
More critically, the study revealed that these cultural tendencies can unknowingly influence users and even permeate society through AI-curated media and educational materials. In other words, generative AI is quietly replicating a certain set of cultural norms and values. Even without directly interacting with language models, we may already find ourselves immersed in the worldview they construct—often without realizing it.
Large language models have become the infrastructure of culture
Language models carry more than just technology—they embody culture. Their built-in corpus and values are gradually shaping our tone standards and modes of communication: the generated response of large language models quietly redefines what tone is “reasonable” and what response is “normal.” In engaging with these models, we are adopting the underlying logic of their value systems.
This is why an increasing number of countries are beginning to incorporate language models into the framework of “sovereign governance.” In 2024, the European Union passed the Artificial Intelligence Act (AI Act), pioneering a risk-based classification system for AI applications. The act also asks that developers of foundation models disclose the sources of their training data, ensuring transparency and oversight of the cultural values embedded within these models.
Singapore, on the other hand, is developing Sea-Lion, an open source large language model for Southeast Asian cultures. Through the extensive collection of Southeast Asian languages and cultural data, Sea-Lion aims to have AI that aligns with local needs and serves as a foundation for building new applications.
Saudi Arabia has made a direct entry through its sovereign fund by establishing Humain, a company led by the crown prince, to develop supercomputing hubs and large-scale data centers, with a total investment of US$100 billion.
All these different strategies are sending the same message: language models are not merely algorithms; they are tied to cultural values, information governance, and national security and must be actively shaped, rather than passively outsourced.
Relying long-term on language models trained overseas means inheriting foreign norms around tone and interaction, and potentially introducing structural risks into our systems. As generative AI continues to permeate industries, if its training data leans toward a particular culture, its values may also seep into society, quietly reshaping how we think and express ourselves.
Japan’s experiment: Shisa.ai — a language model rooted in culture
In the global race for language understanding, Japan has introduced a notably distinctive local initiative: Shisa.ai. This three-person team, with limited resources, successfully trained a 405 billion-parameter Japanese large language model. According to real-world test results, the model performed well in various Japanese tasks, such as prompt comprehension, translation, and semantic reasoning, and is now on par with OpenAI’s GPT-4 and China’s DeepSeek-V3. For a small startup, this marks not only a technical breakthrough but also a profound exercise in cultural agency.
The three founders of Shisa.ai are all immigrants who chose to settle in Japan to start their business. They believe that AI sovereignty must begin with local languages and cultures. Building indigenous models is not only for preserving diversity but also for ensuring data privacy, geopolitical resilience, and national digital autonomy.
Chief executive officer Jia Shen and chief technology officer Leonard Lin are the cofounders of Shisa.ai, which was developed under Leonard’s leadership and is his defining work. One of the AI researchers on the team, Adam Lensenmayer, comes from a distinctly different background. He is a well-known subtitle translator among Japanese animation fans, having worked on titles such as Attack on Titan, Mobile Suit Gundam, Detective Conan (theatrical releases), Galaxy Express 999, Chibi Maruko-chan, Space Battleship Yamato, and even Space Brothers, which can be found in convenience stores across Japan. His meticulous attention to linguistic nuance and tonal precision has made him a key figure in the training of the model. This has allowed Shisa.ai to be more attuned to the subtle contexts and cultural details of the Japanese-language landscape.
From the beginning, Shisa.ai chose to train its models locally in Japan, deliberately focusing on tonal details and social subtext with unique Japanese linguistic features. Founder Jia Shen pointed out that the Internet corpus of the past 30 years has almost been absorbed by large language models. As AI training data approaches its ceiling, future breakthroughs won’t come from “more data,” but “data closer to the context,” such as voices and emotions that are not automatically captured online. From the conversations of the elders and local dialects to the tonal shifts in Gen Z’s dating exchanges—whoever captures the authentic, culturally intertwined corpus of language will hold the key to training the next generation AI models.
This idea is also supported by the Japanese government. In 2024, the Ministry of Economy, Trade, and Industry (METI) launched the Generative AI Accelerator Challenge (GENIAC), which provides funding, mentors, and large-scale computing power to help startups develop their own foundation models. Shisa.ai, one of the selected teams, has successfully accelerated the model-training process and revealed the strength of Japanese language modeling.
What’s more, Shisa.ai has not stopped at technical development. They are actually bringing speech understanding technology to industries, for example, helping Japanese restaurants communicate with international tourists, supporting retail outlets in handling returns and exchanges, and even installing multilingual AI kiosks at train stations to assist travelers with directions and services. These applications are not simply a display of technology but a pragmatic response to cross-cultural communication.
Whose tone of voice will become AI’s voice?
Shisa.ai may be just the starting point, but it reminds us that generative AI is rapidly integrating into everyday life and across industries. The tone it conveys will profoundly shape how we understand one another and express ourselves.
Language is not just a tool, it’s a carrier of culture. The way we speak will ultimately determine how AI understands the world.
Matt Cheng is Founder and General Partner of Cherubic Ventures. Matt is a Taiwanese venture investor, serial entrepreneur, company advisor, and former junior tennis player. Prior to founding Cherubic, Matt co-founded Tian-Ge in China and 91APP in Taiwan, both went public at over $1B+ in market cap. Matt is also a company advisor to Wish and Atomic VC, as well as an early investor in Flexport, Calm, and Hims & Hers.
TNGlobal INSIDER publishes contributions relevant to entrepreneurship and innovation. You may submit your own original or published contributions subject to editorial discretion.
Featured image credit: Cherubic Ventures
The AI economy is here; Are we ready to Identify “AI consumers?”