There is no Australian equivalent of fair use, no text-and-data-mining exception, and no decided Australian case on model training. That combination means the default answer to “can I train on this in Australia?” is no, unless you have a licence — and it means most of what you have read about AI copyright does not apply.
The position, stated plainly
Copyright in Australia is governed by the Copyright Act 1968, available in consolidated form from the Federal Register of Legislation. Reproducing a literary, artistic, musical or dramatic work in material form is an act comprised in the copyright under section 31, and doing it without licence infringes under section 36 unless an exception applies. Building a training corpus reproduces the works in it. There is no provision in the Act permitting reproduction for machine analysis, and none has been enacted.
It is worth being equally plain about what has not happened: no Australian court has ruled that model training infringes, because no such case has been decided. The position described here is the result of reading the statute in the absence of an exception, not the result of litigation. Those are different kinds of confidence, and a page that conflated them would be overstating.
This is a description of the statutory position and not legal advice. Whether a particular corpus, a particular licence or a particular research arrangement is covered turns on facts and on terms this page cannot see. Take Australian advice before relying on any of it.
Why fair dealing does not help
Australia has fair dealing, which is often mistaken for a smaller version of fair use. It is a structurally different thing. Fair dealing applies only where the dealing is for one of a closed list of purposes set out in the Act:
- research or study (sections 40 and 103C);
- criticism or review (sections 41 and 103A);
- parody or satire (sections 41A and 103AA);
- reporting news (sections 42 and 103B);
- giving professional advice by a legal practitioner, patent attorney or trade marks attorney (section 43).
If the dealing is not for one of those purposes, the fairness of it is irrelevant — the enquiry never reaches the fairness question. A court cannot find a new purpose fair, which is precisely what US courts do when they hold a novel use transformative. Commercial model training is not within any listed purpose, and an argument that it is “research” faces both the commercial character of the activity and the way the research provisions have been read as directed at a person’s own research or study.
Why US reasoning does not transfer
This deserves its own section because the error is so common. A great deal of the available commentary on AI and copyright reasons from Bartz v Anthropic or Kadrey v Meta to a general conclusion that training is lawful. Those are decisions of United States district courts applying 17 U.S.C. § 107, an open-ended standard that Australia deliberately does not have. The Australian Law Reform Commission recommended adopting a fair use standard in its 2013 report on copyright in the digital economy and the recommendation was not implemented; the Productivity Commission made a similar recommendation in 2016, with the same result. The absence is a policy choice that has been made repeatedly, not an oversight.
The practical consequence is that a corpus assembled lawfully in the United States on a fair-use analysis may be unlawful to assemble in Australia, and the reverse is also true of Singapore and Japan, whose statutory exceptions are wider than anything Australia offers. Where the copying physically happens is therefore a design decision with legal content, not just an infrastructure one. Compare Singapore’s computational data analysis exception and Japan’s Article 30-4.
The policy review, and its status
Reform has been under active consideration and has not produced legislation. The Attorney-General’s Department established a Copyright and Artificial Intelligence Reference Group in December 2023 as a standing forum for stakeholders, and it has run consultations on transparency and on the training question since; the department’s copyright pages are at ag.gov.au. Separately, the Productivity Commission raised the option of a text-and-data-mining exception in its 2025 interim work on data and digital technology, which drew strong and public opposition from Australian creative-sector organisations.
As at the date on this page, no text-and-data-mining exception has been introduced into the Copyright Act, and no bill doing so is before the Parliament. That is the status: an open policy question with an active consultative process and no legislative outcome. Anything you read describing an Australian TDM exception is describing a proposal, and the distinction between a proposal and a provision is the difference between a plan and a defence. This page is marked as dated for exactly that reason — check the Federal Register of Legislation before relying on it.
What this means in practice
For a team operating in or into Australia, the workable positions are the ordinary ones that do not depend on an exception:
- Licence the data. The only fully clean answer for protected material. Collecting societies and direct publisher licences both exist; the terms need to permit machine analysis explicitly, because a licence to read is not a licence to train.
- Use material that is out of copyright or openly licensed. Check the licence terms rather than the label — several widely used open licences impose conditions, such as attribution or share-alike, that are awkward to satisfy in a model.
- Rely on someone else’s training. Using a commercially provided model does not make you the person who made the training copies. Your exposure shifts to the contract: read the indemnity, its exclusions, and its cap. See AI contract clauses.
- Train where an exception exists, but do it deliberately. Territoriality is a real defence and a fragile one; it needs to be a documented decision about where copies are made, not an accident of where a bucket happens to live.
Australia’s separate work on AI safety regulation is a different track with a different timetable; see the proposed mandatory guardrails.
Top comments (0)