The question of who gets paid when an artificial intelligence system learns from a novel, a song or a news article has become one of the defining legal fights of the decade, and Australia is quietly staking out a position that could prove costly for the technology companies building the models. A new analysis from the University of Melbourne argues that the way Australian law is written leaves AI developers with a clear obligation: if they want to train on protected creative works, they should license them and pay for them.
That conclusion sounds simple, but it lands in the middle of a global argument that is anything but settled. Around the world, courts and parliaments are grappling with whether feeding copyrighted material into a training pipeline counts as ordinary use that requires permission, or whether it falls into some new category that the law never anticipated. The answer carries enormous financial weight, because the value of the world’s leading AI systems rests, in large part, on the vast libraries of human writing, art, music and code they were trained on.
Why Australia’s position is different
The heart of the Melbourne analysis is a point about legal architecture. The United States leans on its flexible “fair use” doctrine, a broad and famously unpredictable standard that AI companies have invoked to defend the mass ingestion of copyrighted material. Several of the biggest cases in the world, including litigation brought by authors, artists and news organisations against the largest model builders, now turn on how far that doctrine stretches. The outcome remains genuinely uncertain, and different American judges have already signalled different instincts.
Australia took a different road years ago. Rather than a sweeping fair use exception, Australian copyright law relies on a narrower set of specific “fair dealing” exceptions that cover defined purposes such as research, criticism, review and news reporting. Commercial-scale copying to train a product that will be sold back to the market does not sit comfortably inside any of those boxes. The practical consequence, the analysis suggests, is that a company scraping Australian creative works to build a commercial model cannot easily point to an exemption that makes the copying lawful without a licence. In other words, the default setting in Australia tilts toward permission and payment rather than toward a free pass.
That framing matters because it reframes a debate that has often been presented as a binary between innovation and obstruction. Under the Melbourne reading, paying creators is not a barrier that Australia has erected against progress. It is simply what the existing law already requires, and the burden of changing that sits with anyone who wants a different outcome.
Two sides that are not close to agreement
Creators and rights holders have been making a version of this argument for some time. Authors, musicians, visual artists, photographers and publishers have watched generative tools produce work that competes with their own, built on foundations that include their labour, and they have asked a straightforward question: if the material has value to the machine, why should the people who made it receive nothing? Collecting societies and industry bodies in the creative sector have pressed governments to confirm that training is a licensable act, and the Melbourne analysis gives that camp a serious piece of legal reasoning to lean on.
The technology industry sees the same facts through a very different lens. Model builders and their backers argue that requiring individual licences for every piece of training data is impractical at the scale modern systems demand, that it would slow the pace of research, and that it risks handing an advantage to jurisdictions with looser rules. Some have gone further, warning that strict licensing regimes could push development offshore entirely, leaving countries such as Australia as consumers of AI rather than builders of it. There is also a competition dimension, because the largest incumbents can afford sweeping licensing deals with major publishers, while smaller local startups may struggle to match them.
Both positions contain real force, and neither is going away. What the Melbourne analysis does is shift the starting point of the negotiation. If the law already requires payment, then the conversation is no longer about whether creators deserve compensation, but about how a workable licensing market can be built so that payment actually flows.
What it means for Australia
For Australia, the stakes are unusually pointed. The country has a deep well of creative output, from a globally recognised screen and music sector to a large publishing and news industry, and much of that material has already been swept into training datasets assembled overseas. If Australian law does indeed require licensing, then a great deal of value that has flowed out of the country without payment could, in principle, be brought back to the people who created it. That is a meaningful proposition for a creative economy that employs hundreds of thousands of Australians and contributes billions to national output.
At the same time, Australia is trying to build a domestic AI industry, and policymakers are acutely aware of the risk of writing rules that make local development uncompetitive. The federal government has been weighing how to encourage sovereign AI capability while protecting the interests of workers and creators, and copyright sits right at the intersection of those goals. A regime that is too permissive hollows out the creative sector; one that is too rigid could choke local model builders before they scale. The Melbourne analysis effectively argues that the existing law has already picked a side, and that the task now is to make it function rather than to relitigate the principle.
There is also a strategic angle for Australian negotiators. If the domestic legal position is that training requires a licence, that becomes leverage in dealings with the multinational firms whose products are already used widely across Australian workplaces, schools and homes. It gives local rights holders and, potentially, the government a stronger seat at the table when the terms of access are set.
What happens next
None of this is fully resolved, and the practical questions are formidable. Who administers licensing at the scale AI training demands? How are payments calculated and distributed fairly across millions of individual works and creators? What happens to models that were trained before any of these questions were answered? Australia will also be watching the international courts closely, because a decisive ruling in the United States or Europe could reshape commercial expectations everywhere, regardless of what local law says on paper.
The likely path forward runs through some combination of collective licensing schemes, industry codes and, quite possibly, legislative clarification as the government continues its broader work on AI regulation. Whatever emerges, the Melbourne analysis has made a useful contribution by cutting through the noise: in Australia, the starting assumption is that the creativity feeding these systems has an owner, and that owner is entitled to be paid.
Sources: The University of Melbourne.
