AI Training Data and Copyright: What Is Clear, What Is Not

Oct 3, 2026

Summary and scope

In Australia, copying copyright works into an AI training set without permission is likely infringement, because there is no US-style fair use and the government has ruled out a text-and-data-mining exception. Overseas courts disagree on whether training itself, or the trained model, counts as copying.

This is a desk review of public sources as at 3 October 2026. It is not legal advice and is not drawn from a case-law database (Westlaw, LexisNexis or similar). Anything relied on should be checked against the primary judgment or statute, and by a lawyer. I found no Australian court judgment on AI training.

Clearly inside copyright

Three acts are clearly within the copyright owner's control, at least in Australia.

• Copying works into a training set. Reproducing a work in digital form is an exclusive right under s 31 of the Copyright Act 1968. Law firm commentary says unauthorised scraping or reproduction of this kind will in most circumstances infringe, and notes Australia has no broad fair use exception.
• Building a dataset from pirated copies. In Bartz, the US court held that training on lawfully acquired books was fair use but storing pirated copies was not. From my own knowledge, that case ended in a very large settlement (not verified in this review).
• Reproducing protected text on demand. In GEMA v OpenAI, a Munich court found OpenAI infringed by using song lyrics in training and outputs, and the court stressed how easily the lyrics could be prompted back out.

Clearly outside copyright

These points come from long-standing copyright principles and my own knowledge, not from this review's searches, so check them before relying on them.

• Ideas, facts, methods and inventions. Copyright protects the way something is expressed, not the underlying idea or information. A patent is a separate right: it restricts making or using an invention, not reading or learning about it. Training on patent documents does not infringe the patent, although the wording of a document can still carry copyright.
• Public-domain works. Works whose copyright has expired can be copied freely. Openly licensed works are usable on the licence's terms.
• General style or technique. Writing in the manner of an author or artist is not infringement unless a substantial part of a specific work is reproduced.
• Material nobody owns. Examples are raw facts and, in Australia, content lacking a human author, since authorship requires a human contribution (the Telstra v Phone Directories line of cases, again from memory).

Contested

Courts have not agreed on whether training is copying, whether a trained model contains copies, and who bears the cost.

Question	One answer	Another answer
Is training on copyright works infringement?	US trial courts in Bartz and Kadrey found generative training transformative and fair use	The Third Circuit on 29 September 2026 rejected fair use for training a competing legal search tool, a non-generative tool built as a substitute for Westlaw
Does the trained model contain copies?	Munich: yes, where works are memorised	UK High Court in Getty v Stability: no, the weights represent learned patterns, not stored works
Does the licensing market count as harm?	The Third Circuit acknowledged an emerging market for licensing training content	Kadrey criticised the plaintiffs' market-harm arguments as half-hearted
Does training abroad escape local law?	UK: Getty's training claim failed because training occurred outside the UK	Munich: liability can attach where a service is offered in Germany

Still open in the US: the Second and Ninth Circuits have not ruled, and cross-motions for summary judgment in NYT v OpenAI were filed on 4 September 2026. The US Justice Department also filed a statement of interest arguing that LLM training is transformative. The UK Getty appeal has been allowed to proceed and may address training itself.

Australia in detail

Australia's defences are narrow, the policy line has been firm, and a leaked September 2026 proposal suggests it may be softening.

What the Act offers a developer

• Reproduction right (s 31). Copying a work into a dataset is reproduction in a material form. Commentary on the Act says the only possible defences are limited fair dealing provisions and narrow temporary-copy exceptions.
• Fair dealing (ss 40 to 43). The five purposes are research or study, criticism or review, parody or satire, news reporting, and judicial proceedings or professional advice. One analysis says research or study is unlikely to cover reproducing entire works for machine learning, and another says a commercial purpose makes it hard to rely on.
• Temporary copies (ss 43A and 43B). These may cover some temporary copies made during learning, but not the copies that build the dataset. A 2026 journal article argues s 43B(2) excludes reproductions of infringing copies altogether, so unlicensed datasets fall outside it.

Policy timeline

Date	Event
Late Oct 2025	Attorney-General Rowland rules out a text-and-data-mining exception after the Productivity Commission floated one, and refers next steps to the Copyright and AI Reference Group, including a possible small claims forum
15 Jul 2026	Prime Minister confirms mandatory national AI standards, an Office of AI and protection of copyright holders
Sep 2026	The ABC reports a leaked consultation paper, "AI on Australian Terms", with two options that would make training opt-out by default, paired with licensing or payment mechanisms

The government has stressed that no changes are finalised, and the Industry Minister has denied a reduction in copyright protection.

Likely test case. I found no Australian judgment on AI training. A test case would probably be brought by a rights holder or collecting society against a developer that copied works while training in Australia, for example on a local data centre. Australian authors and publishers' bodies, and local data-centre growth, make that a live prospect. Training done offshore raises the territoriality problem seen in the UK Getty case. That last step is my inference, not a reported plan.

The policy question

Whether training data should be proprietary is a policy argument, not a legal finding, and Australian law currently answers yes for unlicensed commercial copying. The two sides below are the strongest versions of each case, not my own view. I could not verify the quote attributed to Naomi Klein in the shared meme.

The case that training should be free (or opt-out)

• Reading and learning from public material is how people create, and a model is a form of learning.
• Facts, ideas and style are already unowned, and the model's output is mostly new expression.
• Licensing every source is impractical, and strict rules may push training and investment to countries with looser laws. The leaked Australian proposal is reportedly aimed at attracting AI investment.

The case that training should be licensed

• Training makes commercial copies of works, and the models can substitute for the creators whose work they use, which is the market-replacement worry raised in the Australian debate.
• An opt-out system forces creators to act to protect rights they already hold. Senator Pocock said the burden should not fall on Australians to defend them.
• The Third Circuit recognised an emerging licensing market, which unlicensed training may undercut.

The typewriter analogy and the "theft" framing both overreach. A typewriter keeps nothing of what its operator has read, whereas a trained model can memorise some works. But the law has no general rule that learning from a work needs consent: the question is whether the copying involved is permitted.