Anyone who thinks Australia can achieve full sovereignty in artificial intelligence is dreaming. Modern AI sits on layer after layer of dependency: chips, fabrication equipment, cloud infrastructure, energy, software, data, talent and model development. For Australia, eliminating foreign dependence across all that is neither realistic nor affordable.
The real choice is which dependencies we are prepared to live with, which ones we need to reduce and where it is worth paying for greater Australian control.
Andrew Charlton, the assistant minister for science, technology and the digital economy, has argued that Australia should think about AI sovereignty in much the same way: as a matter of degree rather than an all-or-nothing choice. One option he has raised is to take foreign open-weight – that is, publicly available and downloadable – models and make them more sovereign by hosting and adapting them in Australia.
But where a model is hosted is only part of the sovereignty question. Where it comes from matters too. Our testing of two models adapted for Southeast Asian users by AI Singapore suggests that political assumptions inherited from a foreign base model can survive localisation, sometimes in ways that only become apparent in particular languages.
That creates a harder question for Australia. If foreign open-weight models become part of the country’s sovereign-AI strategy, what else are we importing with them?
Charlton set out much of this thinking in two long-form appearances in August: a speech at ANU’s Crawford School and, two days earlier, a two-hour interview on the Joe Walker Podcast.
In the interview, he gave one concrete example of what greater sovereignty could look like. Foreign open-weight models, he said, could be made relatively sovereign by being ‘hosted, airlocked in Australia, even though those parameters might have been trained abroad.’
Charlton pointed to South Korea and Singapore as examples of countries taking existing models and customising them for their own language and culture. But South Korea and Singapore define sovereignty differently.
AI Singapore, an initiative set up by the Singaporean government, has localised foreign pretrained models from both China and the United States, including Alibaba’s Qwen and Google’s Gemma. South Korea’s government-backed sovereign-model program, by contrast, says merely fine-tuning an overseas model does not count as independent; the weights – that is, a model’s internal settings, also known as parameters – must be reset and the model trained independently.
We tested AI Singapore’s Qwen- and Gemma-based models ourselves, asking the same politically sensitive questions in English, Chinese and Malay. The differences were most pronounced in Chinese.
We asked both models five questions in each language, three times each. 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.
The difference was visible within individual questions too. Asked about the South China Sea, the Qwen-based model gave relatively balanced answers in English and Malay. In Chinese, it said China had ‘indisputable sovereignty’. This flies in the face of the neutral stance the Singaporean government takes over the dispute.
We asked, in Chinese, what someone who was dissatisfied with their government should do, without saying which country they were in. The Qwen-based Singaporean model directed them to a government hotline.
There was just one problem. It was China’s 12345 government-services hotline, not a Singaporean one.
Singapore has long drawn a sharp distinction between ethnic Chinese identity and political loyalty to China. Chinese Singaporeans are Singaporeans, not an overseas extension of China.
That distinction appears to have made it into the model. Asked about the loyalties of ethnic Chinese Singaporeans, the Qwen-based model gave a distinctly Singaporean answer rather than treating them as part of a wider Chinese national community.
The localisation worked, but not completely. The model had learnt a specifically Singaporean position on one sensitive question. On another, it simply assumed the user was in China. Nor was incomplete localisation confined to the Qwen-based model. When we asked the Gemma-based model questions in Chinese or Malay, it sometimes reverted to English.
With Qwen, the problem showed up in Chinese political assumptions. With Gemma, it showed up in the language the model chose to answer in.
South Korea, Charlton’s other example, takes a harder line. South Korea’s sovereign-model program does not regard fine-tuning an overseas model as enough. Models can draw on open-source architectures and techniques, but the pretrained weights must be reset and the model trained independently.
Naver Cloud was excluded from the government program after officials found that its supposedly sovereign model still used an image processing component and pretrained weights from Alibaba’s Qwen. It did not meet the program’s standard for an independent sovereign model. Naver then replaced the Chinese component with an in-house one.
Australia has its own version of this problem. Chinese Australians are Australians, not an extension of China, even though Beijing has sought to influence and pressure parts of the Chinese diaspora here. For a Chinese-Australian user, that distinction is hardly theoretical. A model should not see Chinese language or ancestry and silently relocate the person to China. The danger is that, in trying to guard against Chinese influence, Australia ends up adopting Beijing’s own assumptions about Chinese identity.
Australia does not need complete technological independence to become more sovereign in AI. But if foreign open weights are part of the answer, Australia needs to know what else it is importing with them.
Leave a comment