Technology and privacy · 19 September 2026
In October 2025 the Attorney-General said Australia "will not entertain" a free pass for AI companies to train on copyrighted work. In September 2026 a leaked departmental proposal put an opt-out licence on the table. The difference between those two positions is who controls your work, and who gets paid.
Who should control, and profit from, your work when a machine learns from it?
Australia thought it had answered that question. On 26 October 2025 the Attorney-General, Michelle Rowland, announced that the government "will not entertain a text and data mining exception" to the Copyright Act, adding: "There are no plans to weaken copyright protections when it comes to AI." The Productivity Commission, which had floated exactly such an exception two months earlier, concluded in its final report that it would be "premature to make changes to Australia's copyright laws".
Then, on 15 September 2026, the ABC published leaked documents showing that the Attorney-General's Department had put two options to rights-holder groups earlier in the month. Both would let AI companies train on Australians' creative work. Neither requires the creator's permission first.
The Copyright Act 1968 gives creators the exclusive right to reproduce their work. Australia has no general "fair use" defence of the American kind. It has fair dealing, which is limited to specific purposes such as research or study, criticism or review, parody or satire, and reporting news. Training a commercial AI model on a copyrighted work is not one of them.
There is no text and data mining exception. That is the point the Attorney-General made in October 2025, and the point the leaked proposals would now work around by licence rather than by exception.
The distinction matters. An exception means free use. A licence means paid use. The leaked options are licences, which is why the government can say it has kept its promise. But an opt-out licence means the default is use, and the burden of stopping it falls on the creator. A photographer with ten thousand images, a musician with a back catalogue on three labels, or an author whose publisher is in liquidation is not in a position to opt out of anything.
The sequence is short and worth remembering.
In August 2025 the Productivity Commission's interim report on harnessing data and digital technology proposed a text and data mining exception. Creative industries reacted with fury.
On 26 October 2025 the Attorney-General ruled the exception out and directed consultation on licensing through the Copyright and Artificial Intelligence Reference Group, a standing body the department describes as "a mechanism to facilitate engagement, information sharing and discussion between government and non-government stakeholders". Options canvassed included statutory licensing, collective licensing and voluntary licensing.
In December 2025 the Productivity Commission's final report stepped back, concluding it would be premature to change copyright law and recommending a review of copyright settings over three years. APRA AMCOS chief executive Dean Ormston welcomed the licensing path, saying it "provides the pathway for AI development while ensuring creators are fairly compensated".
In September 2026 the department's two options leaked. Licensing, yes. But on terms that reverse the default.
The strongest evidence of what training data is worth comes from a courtroom in California. In Bartz v Anthropic, a class of authors sued over the use of pirated books to train AI models. A USD 1.5 billion settlement received preliminary approval in September 2025 and final approval on 20 July 2026, working out at roughly USD 3,000 per title across nearly 500,000 books.
Three thousand dollars a book is not a rounding error. It is a market price, arrived at under litigation pressure, for the thing the leaked Australian proposals would license collectively. Any Australian scheme that pays creators materially less than that per work, by default and without their opt-in, will be measured against it.
The United Kingdom offers the other cautionary tale. Its Data (Use and Access) Act 2025 became law on 19 June 2025 only after a prolonged parliamentary fight over AI training and copyright, with creators demanding transparency about what was used to train models. The long title of the Act still carries the scar: it makes provision "about works protected by copyright and the development of artificial intelligence systems".
The leak matters because of what it reveals about process. A licensing model that reverses the default of copyright was being shaped with rights-holder groups in private, and the public found out from a journalist. The Deputy Prime Minister's formulation, historic opportunity plus protection for rights holders, is the balance every government claims. The leaked documents suggest which side of the balance the department drafted first.
Creators are not asking for AI to stop. Most are asking for the thing copyright has always promised: that the person who made the work decides whether it is used, and gets paid when it is. That is a question of control before it is a question of money. And control, under an opt-out licence, is exactly what would change hands.
According to documents obtained by the ABC and reported on 15 September 2026, the Attorney-General's Department put two options to rights-holder groups: extended collective licensing on an opt-out basis, under which AI companies could use works unless the rights holder opted out, or a threshold approach under which an AI company that struck deals with a minimum number of rights holders for a minimum period would gain broader permission.
Yes. On 26 October 2025 Attorney-General Michelle Rowland said the government would not entertain a text and data mining exception and that there were no plans to weaken copyright protections for AI. Consultation on licensing was directed through the Copyright and Artificial Intelligence Reference Group.
The Copyright Act 1968 has no text and data mining exception and no general fair use defence. Fair dealing is limited to specific purposes such as research or study, criticism or review, parody or satire and reporting news. Reproducing a copyrighted work to train a commercial model without a licence is not covered by those purposes.
Its August 2025 interim report proposed a text and data mining exception. Its final report, released in December 2025, concluded it would be premature to change Australia's copyright laws and recommended a review of copyright settings over three years.
In Bartz v Anthropic, a US class action by authors over the use of pirated books in AI training, a USD 1.5 billion settlement received final court approval on 20 July 2026, equating to roughly USD 3,000 per title across nearly 500,000 books.
A scheme under which a collecting society licenses the use of works on behalf of all rights holders in a category, including those who are not its members, usually with a right for individual rights holders to opt out. The leaked Australian proposal would apply this model to AI training.