A leaked Australian government document, presented to the Senate on Tuesday, sets out how AI companies could win the legal right to train on almost anything Australians publish online — photographs, social posts, music, recipes, video — provided they sign enough agreements with the bodies that represent traditional rights holders. Nothing is settled: consultations continue and legislation is expected within a few months. But the direction is readable, and the people the rules reach furthest into are the ones with no organization signing on their behalf.
The document describes several options:
AI companies get the right to use any unprotected material on the internet for training, if they conclude agreements with a sufficient number of organizations and businesses.
AI companies sign agreements with rights-holder organizations and receive broad access to an entire category of content — music, or text — across large numbers of small creators at once.
Companies pay into a central organization that does not yet exist, which distributes the money among "registered" rights holders.
Under any of them, anyone who does not want their material used for training has to defend it by technical means: a password, or a paywall. Both shrink the audience. Owners of their own websites can add a robots.txt file telling bots what they may look at, but rights-holder organizations point out that not every scraper honors the protocol, and material walled off from bots can drop out of search results too. And if the work was posted to a social network, access is controlled by the platform, not by the author.
Lisa Harrison, who lectures in communication and media at Flinders University, argues that technology companies are most likely already collecting material from the web and training new models on it, and that the proposed changes would mainly confirm a legal right to a practice already underway. The reform, in her reading, is catching up with what is happening anyway.
What is at stake for professionals is income. Claire Pullen, chief executive of the Australian Writers' Guild, says creators are already losing commissions to AI that produces imitations of their work and sometimes reproduces it almost verbatim. Industry bodies fear the companies will end up with broad access while creators see no payment. Harrison expects the millions of Australians whose social posts may land in training sets to have even less chance of compensation than the professionals do.
There is a second-order effect that Australian readers, rather than Australian creators, would feel. Foreign artists would gain a reason to shut Australia out. Writers and musicians in other countries have already blocked attempts to use their work for AI training; the UK government abandoned reforms resembling Australia's after Elton John, Paul McCartney and others objected to their work ending up inside models. In the US, Anthropic is defending a multibillion-dollar suit over the alleged theft of "tens of thousands" of songs, and spent US$1.5 billion settling a case over the use of 482,000 books from a pirate database.
Whether the Australian proposals would hand AI companies free access to foreign works distributed in Australia is unclear. Eileen Camilleri, chief executive of the Australian Copyright Council, notes that Australian law protects all content distributed in the country, physically and digitally. Some services already restrict access to television shows from within Australia because their license terms do not permit showing them legally to viewers there. Camilleri says artists, musicians, writers and their publishers could use geo-blocking to close off Australian viewers, readers and listeners if they fear their work being used for training — the digital version of the analog era, when book publishers and record companies controlled distribution territory by territory.
For the companies, the document is candid about the benefit: "legal certainty", plus a solution to the problem of reaching the enormous volume of material held by small creators, where voluntary deals with each one are unrealistic or impossible. The companies would still pay for the large collective agreements. What they would no longer do is negotiate thousands of individual contracts, or absorb litigation costs capable of denting their margins.
Put Anthropic's settlement beside that sentence and the value becomes concrete. US$1.5 billion for 482,000 books is roughly $3,100 a book — a price set by a court process rather than a market, and precisely the kind of number a statutory scheme is designed to make impossible. The reform's real product is not access to Australian content. Australia is a small share of any training corpus. The product is the removal of an exposure.
Pullen, who is in the government-convened negotiations, says the options under discussion resemble the model the AI companies themselves propose. The strongest fact in the whole story is hers: companies can already strike licensing agreements and train models under Australia's existing copyright rules, and while the reform talks continue, not one of them is doing so. That is a revealed preference. Why pay a negotiated rate now for something that may shortly be available at a statutory one?
The companies say a settlement is urgent and the government is expected to introduce a bill within months. Pullen does not see the same urgency among creators and links the haste to the AI companies' coming stock market listings, where they will need to raise as much as possible. That is her inference, not a documented fact, but the calendar fits it better than any argument about Australian data volumes does.
The investment case is the other half of the pitch, and it is the weakest part. Large technology companies call Australia a promising place for AI investment — political stability, a strategic position in the Pacific, available land and natural resources. OpenAI has named copyright as the main obstacle to investing in model training there. Assistant Minister for Technology Andrew Charlton told News24 on Tuesday that Australia has to compete in the US-China race to build high-capability AI and has no ability to "switch off" the process; if the data centers sit offshore, Australian data is stored offshore and the country remains dependent on foreign tokens and foreign AI, which he describes as a vulnerable position.
None of that follows from copyright reform. Training on Australian material does not require hosting it in Australia, and it does not put the country anywhere near the development of frontier models. The question nobody put on the record was answered by silence: on Wednesday, OpenAI and Anthropic both declined to say whether the leaked proposals would be enough to commit them to investing in Australia. On Monday in Canberra, OpenAI vice president Anne O'Leary said the company is in "deep" talks with the government about copyright, and that a company built on human knowledge is the last one that would want that knowledge to stop growing, or people to lose their creative opportunities and the means to create.
So the trade on the table is a definite concession against an indefinite return, with the collecting body that would pay creators still hypothetical and the threshold for "enough" agreements undefined. If it passes in anything like this form, the most visible change for Australians may not be more AI built in Australia. It may be foreign music and video quietly going dark at the border.