"Foreign open-weight models could be made relatively sovereign by being ‘hosted, airlocked in Australia, even though those parameters might have been trained abroad,'" Assistant Minister Andrew Charlton said in a two-hour interview on the Joe Walker Podcast in August.
Andrew Charlton’s sovereignty framing
Andrew Charlton, the assistant minister for science, technology and the digital economy, has argued that AI sovereignty should be treated as a matter of degree, not an absolute. In a speech at ANU’s Crawford School and the Joe Walker Podcast interview in August, he presented one practical option: take foreign open-weight models — publicly available and downloadable — and make them more sovereign by hosting and adapting them domestically. Charlton framed hosting and "airlocking" as a way to gain control even when a model’s parameters were trained overseas.
AI Singapore’s approach: localising Qwen and Gemma
AI Singapore has localised foreign pretrained models from both China and the United States, including Alibaba’s Qwen and Google’s Gemma, adapting them for Southeast Asian users. The article reports testing of two such adapted models across three languages — English, Chinese and Malay — to see how much of the foreign base model’s assumptions survived localisation.
Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleTest findings: Chinese-language political assumptions persisted
The differences showed most strongly in Chinese. The article’s testing found that the Qwen-based model produced responses matching Chinese government positions in 14 of 15 Chinese-language runs; the Gemma-based model did so in two of 15. Examples underline how language choice changed political framing:
- Asked about the South China Sea, the Qwen-based model gave relatively balanced answers in English and Malay but, in Chinese, stated China had "indisputable sovereignty."
- Asked in Chinese what someone dissatisfied with their government should do (no country specified), the Qwen-based model directed the user to a government hotline — specifically China’s 12345 government-services hotline — rather than a Singaporean contact.
- On the question of ethnic Chinese Singaporeans’ loyalties, the Qwen-based model gave a distinctly Singaporean answer, reflecting a separation between ethnic identity and political loyalty that Singapore maintains.
- The Gemma-based model sometimes reverted to English when asked in Chinese or Malay, showing incomplete localisation in the choice of response language rather than in political content.
South Korea’s standard and the Naver Cloud example
South Korea’s sovereign-model program takes a stricter position: it does not regard fine-tuning an overseas model as sufficient. The program requires that pretrained weights — a model’s internal settings, also known as parameters — be reset and the model trained independently. The program excluded Naver Cloud after officials found its reportedly sovereign model still used an image-processing component and pretrained weights from Alibaba’s Qwen. Naver then replaced the Chinese component with an in-house one to meet the program’s standard.
What this means for technologists, policymakers, and Chinese Australians
- Technologists and security teams: If Australia follows Charlton’s "host and airlock" route, teams will need to test not only where a model is hosted but what assumptions survive localisation — language-specific behaviours, default contact information, and geopolitical framings.
- Policymakers and procurement leaders: The South Korean example demonstrates a spectrum of standards. Policymakers must choose whether to accept fine-tuning and hosting as sufficient or to demand deeper re-training that resets pretrained weights, as South Korea requires.
- Chinese Australians and community leaders: The article highlights a concrete risk: a model that treats Chinese language or ancestry as evidence the user is located in China. For Chinese Australians, who are citizens, not an extension of another state, that silent relocation of identity or location has real implications for trust and safety.
Modern AI, the article notes, rests on layers of dependency — chips, cloud infrastructure, data, talent and model development — and Australia cannot realistically eliminate all foreign dependence. The trade-off is a policy choice: which dependencies to accept, which to reduce, and where to pay for greater control. Crucially, the article concludes, if foreign open weights become part of Australia’s sovereign-AI strategy, Canberra must reckon with what else is being imported alongside those weights.
