On 15 July Anthony Albanese returned to his alma mater, the University of Sydney, to give a speech titled AI in Australia’s interests. “Not everything produced in Australia is up for grabs,” he warned the technology sector. “Australian writers, musicians, artists and journalists must retain ownership and control of their work.” No company, he said, should use Australian books, music, art or news to build or train AI without the artist’s control, and that control must extend to “the price and value of their work”. “Anything less,” he said, “is theft.”
The same day the government established an Office of AI inside the Department of the Prime Minister and Cabinet, charged with designing a set of Australian Standards for AI. The standards will set mandatory rules for large data centres, including where they may be built and the power and water they may draw, and the office’s brief says they will come “with strong copyright protections for Australian creators”. Premiers will take up the scheme at National Cabinet later this month. The Prime Minister has said he hopes to release an exposure draft at the end of the year, followed by a bill early in 2027.
Australian writers, musicians, artists and journalists must retain ownership and control of their work.The Hon Anthony Albanese MP, Prime Minister of Australia
The speech confirmed an earlier decision announced by the Attorney-General, Michelle Rowland, on 26 October 2025, when she ruled out the text and data mining exception the Productivity Commission had floated two months earlier. There will be no exception. AI developers who want Australian songs and recordings will need a licence, and the Prime Minister has said the law will “spell that out, plain as day”. The primary question this poses for music rights holders is what a license will cost and who will negotiate it on their behalf. Those who proactively put their ownership records, contracts, metadata and licensing mandates in order over the coming months will be best place to address both.
The relevance of data centres
A Treasury briefing released under freedom-of-information laws, reported by the ABC on 14 July, records Anthropic, the developer of Claude, telling the government that major Australian investment would be “contingent on clarity of copyright settings”. The Australian Financial Review reported on 5 July that Anthropic wants at least 1.4 gigawatts of Australian data centre capacity, at a reported cost of up to US$15 billion. Anthropic accepts that copyright exists, and its complaint concerns clearance, since deals can be struck with the major labels and publishers but licensing material from thousands of smaller rights holders is less appealing to larger firms in the sector. The Australian Recording Industry Association (ARIA) said in June that around 80% of the world’s sound recordings could be licensed for AI training with “four deals”, one with each major label and one with Merlin, and after the Prime Minister’s speech its chief executive, Annabelle Herd, urged AI companies to “pick up the phone and start licensing”.
The relevance of data centres in such a discussion stems from the territorial nature of Australian copyright law. Sections 36 and 101 of the Copyright Act 1968 (Cth) catch a person who, without the owner’s licence, does in Australia any act comprised in the copyright, or authorises it here. A model trained on servers overseas is therefore harder to pursue for copyright infringement under Australian law, particularly for independent and emerging artists and labels. Conversely, model trained in a data centre in Sydney is caught within the scope of ss 36 and 101. That is why the promise of US$15 billion in Australian compute is, to some extent, contingent on the “clarity” of copyright law for AI companies, and why onshore training strengthens the hand of Australian rights holders. Kathy Bowrey of the University of New South Wales told the ABC it is “probable” that some AI training has already infringed copyright, but no AI copyright case has yet reached an Australian court.
Australian copyright law now
Australia has no general fair use defence of the kind found in s 107 of the United States Copyright Act. The Copyright Act 1968 (Cth) allows fair dealing only for specific listed purposes. For musical and literary works those are research or study (s 40), criticism or review (s 41), parody or satire (s 41A), reporting news (s 42) and professional advice (s 43), and ss 103A to 103C apply the first four to sound recordings. Training a commercial model on songs and recordings does not fit easily within the scope those listed purposes, and a use outside the lists and outside any other specific exception may give rise to infringement. Successive governments have declined to import fair use since the Australian Law Reform Commission recommended it in 2013, and the National AI Plan of December 2025 confirms there will be no broad text and data mining exception either.
Overseas in the United States, where fair use is available, Anthropic still paid US$1.5 billion to settle claims over pirated books, being about US$3,000 for each of some 482,000 works, in a settlement given final approval on 20 July 2026. Universal settled with Udio in October 2025 and Warner settled with Suno in November 2025, in each case with licensed platforms due in 2026, and Warner’s deal gives artists and songwriters control over how their names, images, voices and compositions are used. Warner, Merlin, Kobalt and the National Music Publishers’ Association have since signed their own Udio licences, while Universal and Sony’s case against Suno remains ongoing. In Germany, the Munich Regional Court held on 11 November 2025 that lyrics memorised inside OpenAI’s models were reproductions of GEMA members’ works, and on 31 July 2026 it held Suno liable on the same basis and ordered it to stop and to disclose its revenues. Suno says it may appeal. Licences for AI training are being written and priced now, and developers who train without one are being made to answer for it in court.
You knew that your IP was being trained on, but then to see it in the dataset … it really becomes real how much of your work exists within the training data of these models,George Nicholas, Grammy-nominated Australian mixing engineer
Forthcoming
The Attorney-General’s Department consulted its Copyright and Artificial Intelligence Reference Group between October and December 2025 on three models for licensing AI training, and on two further questions: how copyright should treat AI-generated output and the case for a small claims forum for modest infringements.
Voluntary licensing would leave the parties to their own devices, which is largely the status quo. A statutory licence would compel access at a price set by agreement or by the Copyright Tribunal, with no right to refuse. Extended collective licensing would let a collecting society such as APRA AMCOS or PPCA license an entire repertoire, non-members included, subject to a right to opt out. The department has said it will publish the paper and a summary of feedback, and it has not yet indicated a bias toward one of these approaches. The Media, Entertainment and Arts Alliance wants any scheme to give the people who wrote and performed the work a right to payment that cannot be contracted away. Who is covered, who may opt out, how use is measured and how money is split between writers, performers, labels and publishers are all unconcluded, and those are the terms rights holders can still influence.
The second process is the Office of AI’s standard. What has been announced in definitive terms concerns large data centres. The copyright element is a commitment, in the Prime Minister’s words, that “our laws will spell that out, plain as day”, with the particular model to follow consultation. In the meantime, training in Australia on unlicensed music infringes its copyright now.
What music rights holders can do now
The Prime Minister has committed the government to a market in which the owner of a song or recording sets the permission and the price, and licensing income from AI may become easier to collect once a scheme exists. There is plenty for rights holders to do in the meantime.
1) Documentation of ownership
Copyright in Australia arises automatically when a song is written down or recorded. There is no public register as with other forms of intellectual property that lists ownership. Where copyright has been assigned, it must be in writing and signed by the person assigning away their rights, particularly in the case of producer, session musician, co-writer and remix agreements. The same applies to songwriting splits, which should be agreed, documented in writing and signed. Every recording should have an ISRC, available through ARIA, and every work an ISWC through APRA AMCOS. Works should also be registered with APRA AMCOS and recordings with PPCA. Any licensing scheme, society distribution or dataset search will identify the musical works through those identifiers, and as with many assets, the value of a catalogue with a coherent and clear chain of title has greater value to an AI licensee or a buyer than one without.
2) Existing agreements
Older recording, publishing, production and distribution agreements often grant rights in all media “now known or hereafter devised”. No Australian court has yet determined whether such words include the training of an AI model. The practical implication applies to who can license training on your recordings and compositions without asking you. Artists and writers should find out what their label, publisher, distributor and any aggregator can do under their current terms. Labels and publishers should check the same for their distributors and sub-publishers, and should note that distributor policies are adapting. DistroKid now asks uploaders to disclose AI involvement, and Believe and TuneCore refuse tracks made with AI generators that hold no licence.
3) AI terms in agreements
New recording, publishing, producer, session, distribution and sync agreements should confirm whether AI training uses are granted or reserved, and if granted, at what price and with what approvals and reporting. Artists and writers should ask for approval rights over AI uses of their voice and their songs, and for a share of any AI licensing income. Labels and publishers should consider which AI uses they will license, on what terms, and how they will account to artists and writers, since the MEAA is pressing for an unwaivable share by statute and any scheme is likely to be judged on how income passes through it.
4) Who licenses on your behalf?
The major labels and the larger independents are doing direct deals. Most songwriters and independent labels are likely to license collectively through APRA AMCOS and PPCA, or through an intermediary such as Merlin, Kobalt or a distributor with its own AI deals. Whichever approach the government adopts, the mandate you have given, or withheld, will inform who negotiates and the direction of revenue earned. Independent labels and publishers should form a view on the three licensing models and put it to the department and the Office of AI prior to the exposure draft. Artists and writers should ask their societies and their labels what they intend to do with the AI rights they administer.
5) Reserving rights
State on your website and in your distributor metadata that AI training is not licensed, and use machine-readable reservations such as the W3C’s TDMRep protocol and robots.txt. In Australia these create no new right, since there is no exception to opt out of, but they remove any argument of implied licence, and from 2 August 2026 the European Commission can fine general-purpose AI providers under the AI Act for failing to honour such reservations, wherever the training took place. Then look for what has already been taken. The Atlantic’s AI Watchdog database, searchable by artist, shows songs from more than 128,000 APRA AMCOS members in four AI training datasets. A listing in that database suggests your song may have been collected, short of proof that a model was trained on it, so it is worth recording any searches with dates and screenshots, keep copies of unlicensed outputs that copy your work, and waching for claims processes in overseas settlements.
6) Voice and identity
Australian law presently gives no one a property right in their voice or likeness, and a cloned voice is not a copyright work. Protection of this may be found under the law of passing off and s 18 of the Australian Consumer Law where a use suggests your endorsement, from moral rights of attribution and integrity, which extend to performers, from trade marks in stage names and from contract. Artists should consider registering their stage name as a trade mark and clarify the voice and likeness clauses in any deal, and managers may wish to use platform impersonation policies, such as Spotify’s, as a first takedown route. The United Kingdom has said it will consider a new right over digital replicas and Australia has not proposed one, though the bill expected in 2027 would be an appropriate place to do so.
7) Use and disclosure of licensed tools
If you or your artists use generative tools, it is important to use ones that hold licences and disclose the use through the DDEX AI credits that platforms such as Spotify now display. Outputs from unlicensed generators may be refused by distributors, may reproduce someone else’s work, as the Munich court found with Suno, and may attract no copyright at all, since Australian courts have required a human author for a work to be protected. Until the consultation on AI outputs changes that, it is important to keep a record of the human contribution to anything you release.
National Cabinet is expected to meet later this month, the Attorney-General’s Department has yet to name its preferred licensing model, and the bill is promised for early 2027. Mr Albanese has told the AI sector that Australian culture is available on Australian terms, and Anthropic’s general counsel, Jeffrey Bleich, says the company takes seriously its “responsibility to meet the terms set out by the Australian government for AI developers”. Australian rights holders are best prepared when they have sound documentation of what they own and have decided who may license it.
If you would like advice on what these developments mean for your catalogue or your contracts, we would be glad to hear from you.
This article is general information, current at 21 August 2026. It is not legal advice, and must not be relied upon as such. No lawyer-client relationship is created by reading this article.
