Generative AI asks copyright law two questions it was never designed to answer. Who owns what comes out of the model? And who is owed what for the material that went in?

As at November 2025, Australia’s answer to the first is an accident of doctrine written prior to anyone having heard of a large language model. Australia’s answer to the second was settled, at least for now, just last month, when the Government publicly killed off the idea of a text and data mining exception.

Question one: who owns the output?

Australian copyright law has no provision dealing with AI-generated works, but it does not need one to reach a result. The Copyright Act 1968 (Cth) protects original works, and two decades of authority establish that originality is inseparable from human authorship.

The key relevant cases below predate generative AI entirely:

Case	What was generated	Why it mattered
IceTV v Nine Network (HCA, 2009)	TV programme schedules	Copyright protects the independent intellectual effort of human authors, refocusing the law on authorship rather than industrious collection
Telstra v Phone Directories (FCAFC, 2010)	Phone directories compiled by automated systems	No identifiable human authors exercising independent intellectual effort, so no copyright, however valuable the product
Acohs v Ucorp (FCAFC, 2012)	HTML source code produced automatically by software	Code written by a program rather than a person is not an original work of authorship

Applying that line of authority to a Midjourney image or a ChatGPT-drafted document, the likely outcome is stark: purely AI-generated content attracts no copyright in Australia at all. Not ownership by the AI company, not ownership by the user. A void. Anyone could copy such material freely, because there is nothing to infringe.

The patent register tells the same story in a different register: in Thaler v Commissioner of Patents (2022), the Full Federal Court held that only a natural person can be an inventor.

The commercially important question sits at the boundary. Almost nobody uses these tools with zero human input. Somewhere between typing a two-word prompt and substantially rewriting a machine draft, a human contributes enough independent intellectual effort to become an author of the result. No Australian court has yet decided where that line falls, and until one does, every business generating logos, marketing copy, product descriptions or code with AI assistance is holding assets of genuinely uncertain ownership.

Question two: who pays for the input?

The training question bites harder in Australia than the headlines from overseas suggest, for a structural reason. Australia has no general fair use defence. It has fair dealing, a closed list of permitted purposes: research or study, criticism or review, parody or satire, news reporting and a handful of others. Copying millions of works to train a commercial model does not fit comfortably within any of them. Whatever room foreign developers may have under their home regimes, training on Australian works within Australia sits in a legally exposed position by default.

That default is exactly what the Productivity Commission proposed to change, and what the Government has now declined to change. The arc ran quickly:

• July 2025: the Tech Council’s Scott Farquhar uses a National Press Club address to urge the Attorney General to amend the Copyright Act to accommodate text and data mining, framing it as unlocking billions in foreign investment
• 5 August 2025: the Productivity Commission’s interim report, Harnessing Data and Digital Technology, floats a text and data mining exception among its options, estimating AI could add around 116 billion dollars to GDP over a decade. The Commission stresses the exception would be conditioned on fairness and would not be a blank cheque
• August to October 2025: the creative sector mobilises. Music bodies APRA AMCOS and NATSIMO, the Australian Society of Authors, the Copyright Agency and the Australian Copyright Council all oppose the proposal, arguing it would legitimise the unlicensed copying that has already occurred and hand Australian cultural output to foreign developers for free
• 26 October 2025: Attorney General Michelle Rowland announces the Government will not introduce a text and data mining exception, framing the decision around certainty for creators and fair compensation, and directs further work on licensing through the Copyright and AI Reference Group

The rejection is a genuine policy choice, not a deferral. Several comparable jurisdictions have adopted training-friendly exceptions in some form; Australia has deliberately declined to follow. But it is worth being precise about what the decision does and does not do. It does not make training on Australian works lawful or unlawful. It simply preserves a status quo in which such training, without a licence, is presumptively infringing and no bespoke defence exists.

The unresolved work, now handed to the Reference Group, is how a licensing market for training data might actually function: collective licensing through existing societies, direct deals between developers and publishers, or something built for purpose. None of that existed as at November 2025.

Possible directions

Neither question will stay unanswered forever, and the contours of the reform debate are already visible.

On ownership of outputs, the courts have reaffirmed the human-centred character of intellectual property law, but policy discussions about reform are underway in many jurisdictions, and the options tend to cluster into four camps:

• Keep the current approach and treat AI as a tool. Protection would turn on demonstrating sufficient human involvement in the creative or inventive process, much as photography was eventually accommodated as a tool of human authorship. This is the least disruptive path, but it leaves the hard boundary-drawing to the courts on a case-by-case basis.
• Attribute authorship or inventorship to the human developer or operator of the system. This recognises the role of those who design, train or direct the technology, and resembles the deeming approach some jurisdictions already take for computer-generated works, though it strains the traditional link between authorship and creative contribution.
• Create a new, purpose-built form of protection for AI-generated works. A sui generis regime, likely shorter in duration and narrower in scope than copyright, would recognise the distinctive character of machine-generated output while avoiding the conceptual difficulty of treating an AI system as a legal person.
• Leave AI-generated works unprotected in the public domain. On this view, machine-generated material should be freely available to society, with exclusive rights reserved as an incentive for human creativity. It is the current Australian default by accident; some commentators argue it should be the position by design.

On payment for inputs, the live discussion following the text and data mining rejection is whether Australia should move from purely voluntary licensing toward a statutory licence framework. Under such a scheme, developers could train on Australian works without individual permission but would owe remuneration, most plausibly collected and distributed through existing collecting societies, echoing the statutory licences that already operate for education and government use. The attraction is that it solves the transaction-cost problem of licensing millions of works one by one; the objection, pressed by parts of the creative sector, is that a compulsory scheme still takes choice away from creators and merely prices the taking. Where the Copyright and AI Reference Group lands between voluntary deals, collective licensing and statutory compulsion is now the central question in Australian AI copyright policy.

Australia has decided whose rights it will not take away. It has not yet decided who holds the rights to what the machines make, or how those whose work feeds the machines will be paid.

The next chapter of this debate will be written in exactly those two gaps.