U.S.-based artificial intelligence (AI) company Anthropic received an “export control” directive from its government. The company was told it must block access to two of its most capable models, Fable and Mythos, for all foreign nationals.
Within hours, Anthropic shut down access to the models for users everywhere in the world, including researchers, clinicians, and analysts in Australia. This happened with no warning and no backup plan.
Why did this happen?
The directive’s “foreign national” criterion is a citizenship concept. However, Anthropic and other cloud-based AI providers only know the location of their users, not their citizenship.
Consumer AI services have no effective mechanism to verify citizenship. Even their location filters can be dodged with tools such as virtual private networks (VPNs).
There is no way that any control of these services based on user nationality can be enforced. A system that blocks a foreign national in the United States, but allows access to a U.S. national in Australia, for example, is not currently available.
The U.S. government directive was a geopolitical signal — an indication rather than an enforceable control. The only way Anthropic could comply was to shut down access for people everywhere.
A question of sovereignty
Australian universities, government agencies, health systems, and industry have integrated U.S.-hosted frontier AI deeply into their operations. Advanced analytics platforms, AI-assisted research tools, and productivity infrastructure built on top of models such as Claude or GPT-5 operate with an implicit assumption: that access will persist.
The same is true of ubiquitous systems such as Microsoft Teams or Google Workspace, but as we are seeing, the assumption may not hold for AI.
The Anthropic shutdown was a reminder that access and control sat entirely within U.S. jurisdiction, regardless of where any Australian user happened to be. It represents a failure of data and technology sovereignty: our ability to operate without permission from other nations.
Australia’s exposure
The Anthropic shutdown is the first legally enforceable export control targeting a software-level AI system. It will not be the last.
Any frontier AI model hosted in the U.S. by Anthropic, OpenAI, Google DeepMind, or Meta is subject to U.S. export control laws. The precedent set this week extends to all of them.
There is some debate about whether an export control only applies to physical exports and not remote access to models housed in the U.S., and some experts have suggested the order may be challenged on those grounds. If that were to happen, we would expect the U.S. government to change the regulations.
The U.S. government has previously used technology export controls as a geopolitical instrument, particularly in the case of chips and semiconductors. Even U.S. allies such as Australia are not exempt from these controls.
The Anthropic order should have come as no surprise. Voices in Australia’s own research community, including Jon Whittle, former head of CSIRO Data61, had been publicly warning about exactly this scenario for more than a year.
To prepare for the future, Australia needs a strategy for its own sovereign AI. This can’t be a distant aspiration: it needs to be an operational plan with named owners, timelines, and a budget.
How does Australia achieve AI sovereignty?
A true sovereign AI strategy requires four things. First is data sovereignty: data that is physically stored in Australia and subject to Australian law. Second is compute sovereignty: in-country data centers under Australian control. Third, AI model sovereignty: an AI capability that is not dependent on a foreign provider. And finally, policy sovereignty: the ability to set its own rules, rather than inheriting another country’s export controls or safety regulations by default.
Australia already has some examples of sovereign data and computing, such as Macquarie Government, Vault Cloud, and AUCyber. Building this capacity will require huge, sovereign data centers on Australian soil.
However, that build-out will only garner public support if it comes with binding commitments on renewable energy sourcing, water-efficient cooling, and transparent community consultation in regions where data centers are proposed (such as Bundey in South Australia).
For Australian residents, this sovereignty means not depending on AI tools and systems that another nation can switch off, restrict, or alter at will.
Developing sovereign AI models will be a challenge. However, it doesn’t necessarily mean building new systems from scratch.
Switzerland’s Apertus model, released in September 2025, is an excellent example of what open-source approaches can produce. It also shows how large language models can be made to comply with regulations such as the European Union’s Artificial Intelligence Act.
Beyond this, sovereign capability will also require training future experts. We can expect increased demand for people who can build the technology, deploy and manage it, govern it, and integrate it into existing systems.
The regulatory dimension
The EU is developing AI policy frameworks from which Australia can learn. It is establishing reciprocal access rights with technology partners, doing institutional contingency planning, and putting money into the process.
Several European countries are moving rapidly. France has just announced a further €655 million (A$1 billion) going into national AI capabilities, explicitly referencing last week’s events. This builds on roughly €2.5 billion of investment since 2018.
Sovereign AI does not mean reinventing the frontier model from scratch. It can mean public, transparent, nationally governed infrastructure built on open foundations.
At a minimum, AI sovereignty requires making sure we are not too dependent on single suppliers or even suppliers based in a single jurisdiction. The goal should be that no single government or jurisdiction can unilaterally cut off Australian residents and institutions.
Bernadette Hyland-Wood, Research Fellow, Digital Media Research Centre, Queensland University of Technology; Dimitri Perrin, Head of School (Computer Science) and Co-Director (Center for Data Science), Queensland University of Technology, and Michael Guihot, Associate Professor, Queensland University of Technology
This article is republished from The Conversation under a Creative Commons license. Read the original article.
Follow us on X, Facebook, or Pinterest
