For much of the past two years, the Australian conversation about artificial intelligence has fixated on where the machines will live. Data centres, power grids and water use have dominated the headlines, and the Albanese government has leaned into the pitch that Australia can be a trusted, stable home for the enormous compute clusters that train modern AI. What has received far less attention is a quieter question that may matter just as much: not where the models are housed, but where they are actually taught.
A new analysis from the Australian Strategic Policy Institute argues that global AI companies are now scoping Australia specifically as a place to run frontier training, the resource-hungry process of building the largest and most capable models, outside the United States. According to the piece published on ASPI’s The Strategist, these firms are broadly willing to meet the government’s expectations on data centres, from energy to security. The obstacle is not steel and silicon. It is copyright, and the analysis frames it as a time-sensitive blocker rather than a background irritant.
Why training is the harder problem
Hosting an AI service and training one are different beasts. Hosting means running a finished model so that Australians can use it, which is largely a matter of servers, connectivity and demand. Training means feeding a model vast quantities of text, images, audio and code so it can learn patterns in the first place, and that is where copyright law bites. Much of the material that makes a frontier model useful is the creative and journalistic output of real people, and under Australian law that output is protected by default.
The United States has muddled through this tension using its flexible fair use doctrine, which courts have been asked to stretch around AI training with mixed and still-contested results. Australia has no equivalent. Our copyright framework relies on a narrower set of fair dealing exceptions for specific purposes such as research, criticism or news reporting, and none of them was written with the wholesale ingestion of a nation’s creative catalogue in mind. For a company deciding where to build its next training run, that legal uncertainty is a genuine variable, not a technicality.
Two camps, one unresolved fight
The debate breaks cleanly into two viewpoints that have been circling each other for more than a year. On one side sit the developers and parts of the tech industry, who argue that without a workable pathway to train on Australian data, the country will simply be skipped. In their telling, the models will be trained elsewhere, Australia will import the finished product, and any ambition to build sovereign capability or capture high-value jobs will quietly evaporate. They point to the risk that overly rigid rules push activity offshore while doing nothing to actually protect local creators, since the same models get trained on the same material in another jurisdiction anyway.
On the other side are authors, musicians, visual artists, news publishers and the unions that represent them, who see an existential threat dressed up as progress. Their position, aired repeatedly through the government’s own consultations, is that letting AI firms train on creative work without permission or payment amounts to the largest uncompensated transfer of value the cultural sector has ever faced. Australian artists have been vocal about wanting a genuine seat at the table rather than a fait accompli, a theme FluentSea has covered as the federal AI office has taken shape.
The ASPI analysis tries to thread this needle by reframing the question. Instead of treating creator compensation and AI training as a zero-sum trade-off, it argues the two can be unlocked together, with a mechanism that pays rights holders while giving developers the legal certainty to train onshore. The detail of any such scheme is where the difficulty lives, because a licensing or remuneration model has to be broad enough to cover the scale of frontier training yet fair enough that individual creators actually see money, not just a headline settlement between big platforms and big publishers.
The Australian stakes
For Australia, the stakes run in two directions at once. There is an economic prize, in the form of investment, sovereign AI capability and the kind of high-skill work that clusters around training infrastructure rather than mere hosting. There is also a cultural cost to get wrong, because a poorly designed exception could hollow out the value of Australian creative work at precisely the moment the country is trying to grow its cultural exports.
The policy machinery is already turning. The Attorney-General’s department has convened a Copyright and AI Reference Group to work through exactly these questions, and the government’s broader AI agenda has repeatedly bumped up against the copyright issue, including a well-documented split within Labor over how far to go. The Productivity Commission has also waded in, floating the idea of a text and data mining arrangement that would ease training while, in theory, preserving creator interests. None of these processes has yet produced a settled answer, which is precisely the gap the ASPI piece is warning about.
The time-sensitive framing matters here. Frontier training decisions are being made now, on the basis of where the electricity, the chips and the legal clarity line up. If Australia spends another two years consulting while the United States, the United Kingdom and parts of Asia offer developers more predictable rules, the analysis suggests the window to be a training destination rather than just a hosting one could close before the country has decided what it wants.
What happens next
The immediate test is whether the government can produce a copyright position that is specific rather than aspirational. That means naming a mechanism, whether it is a collective licensing scheme, a statutory remuneration right, an opt-out register or some hybrid, and putting numbers and timelines around it. Creators will judge any proposal on whether it delivers real payment and real consent. Developers will judge it on whether it removes enough legal risk to justify pointing a training run at Australia.
The uncomfortable truth threaded through the ASPI analysis is that doing nothing is also a choice, and not a neutral one. Leave the law ambiguous and the likely outcome is the worst of both worlds: training happens offshore, Australian work still ends up in the models via other jurisdictions, and local creators are compensated by no one. Getting a workable deal that pays creators while unlocking onshore training would be a genuinely Australian answer to a problem the rest of the world is also fumbling. The question is whether the policy can move at the speed the technology already is.
Sources: ASPI Strategist.
