AI companies are scoping Australia as a country for frontier AI training outside the US. They are ready to meet the government’s data centre expectations, but copyright is a time-sensitive blocker as they are not on track to reach an agreement with rights holders. To discourage AI companies from turning to viable alternative countries, Canberra should provide a solution that’s also beneficial for Australian creators, navigates political red lines and allows Australia to secure its national interests.
One option is an Australian AI training scheme, proposed by Good Ancestors, of which I am chief executive. Consisting of two separate components – a permit for AI training and a fund to support creators – the scheme would unlock AI training, avoid a text-and-data-mining (TDM) exception and deliver more for Australian creators than the likely alternative options.
The AI training permit would be available to entities conducting AI training in Australia. It would require a fee paid into consolidated revenue, an opt-out for Australian rights holders, compliance with the data centre expectations and other conditions of national interest. In return, companies would get legal certainty that their training is not subject to copyright restrictions, and would receive government assistance to accelerate ‘time to power’ for new infrastructure.
An Australian AI public benefit fund, modelled on the Public Lending Right Act 1985, would support creators and rights holders with flexibility in quantum and design. The Public Lending Right scheme is a long-standing mechanism that compensates Australian authors and publishers for loss of sales from their books being available in public libraries. The model could be readily adapted for a broader pool of creators and rights holders.
The permit and the fund would be decoupled. This avoids Britain’s problem of aligning a technical question (what’s inside a dataset) with legal and practical ones (how to opt out and who gets paid).
This is the opposite of a TDM. A TDM gives AI companies a free pass: the right to train without obligation. Instead, the permit extracts maximum value from AI companies based on what they value most, access to the Australian market.
AI training uses data or works that may be protected by copyright. Licensing that material is the obvious response, but on its own it isn’t a realistic path to AI training in Australia.
There are two reasons for this. First, there is no existing licence or combination of licences that gives enough certainty to enable large-scale AI training in Australia. Second, AI training involves global content – more than Australian licences or reciprocal agreements cover. A company training AI in Australia could sign every available licensing agreement and still face litigation from the rights holders it couldn’t identify or reach. Conversely, AI trained here on only readily licensable data would not be competitive at the frontier.
The scheme also preserves valuable licensing opportunities. A permit would not authorise circumvention of technological protection measures (TPMs), such as paywalls. The commercial deals AI companies have struck with companies like HarperCollins and the Associated Press involve content behind TPMs. Those deals remain on the table. The same is true for private datasets or proprietary scientific data.
AI companies have options outside Australia. They could double down on the US. While litigation is ongoing, early rulings have favoured AI training as fair use, and the Trump administration is not going to kill the goose that lays the golden egg. AI companies could also secure pathways in other desirable countries like Canada or Japan, which already has AI-friendly laws and scalable nuclear power.
The Australian government has adapted copyright before, from the photocopier to the internet. Each time, the approach was to enable the technology while securing a fair return for creators. An Australian AI training scheme sits in that tradition. It aims to deliver what rights holders are asking for – consent-control and compensation – for more money and across a broader pool of creators than direct licensing is likely to achieve.
AI training creates a different kind of value than the AI inference investment Australia is currently receiving. Training is building new AI models, while inference is running existing models. Strategically, training is more like a fuel refinery, while inference is more like a petrol station. Both have value, but only one gives you a critical place in the value chain.
Losing frontier AI companies to other countries would be harsh for Australia. If they choose to build training data centres in Canada, Japan or elsewhere, Australia loses its negotiating position. Creators and rights holders would remain restricted by the ‘fair use’ provisions or TDMs in other countries, with little or no compensation and no ability to opt out. Australia would lose the economic benefits of AI training and the trusted relationships needed to ensure access to frontier AI models, which are increasingly critical to national security.
An AI training scheme could put Australia in a small group of countries hosting frontier AI training, with the accompanying economic and strategic benefits. Creators and rights holders would receive more compensation and stronger control over their work than direct negotiation is on track to deliver. The government would retain ongoing levers to adjust the permit or fund as technology evolves.
The window is narrow, but a win-win solution is on the table. Creators, rights holders, technology companies and others are invited to share their views on the proposal – to strengthen it, or to find better alternatives.