The idea that a machine could learn to write a passable pub-rock anthem by quietly ingesting the back catalogues of Jimmy Barnes, John Farnham and Kylie Minogue once sounded like a thought experiment. In 2026 it is closer to a business model. Generative music tools trained on vast libraries of recorded sound are churning out original-sounding tracks on demand, and a growing chorus of Australian artists and rights holders say their work has been swept into the training data without consent, credit or a cent in payment.
A pointed piece in Startup Daily put the problem in blunt terms: some of the country’s most recognisable performers are being fed into AI systems, and there is not a great deal Kylie, Farnsy or Barnesy can actually do about it under the law as it currently stands. The headline was cheeky, but the underlying question is deadly serious for an industry that already runs on thin margins.
How the music ends up in the machine
Modern generative audio systems learn by analysing enormous quantities of existing recordings, absorbing the statistical patterns that make a chorus feel like a chorus or a guitar tone sit a certain way in a mix. The developers behind these tools generally do not license each track. They scrape or otherwise acquire large datasets, run them through training, and argue that what comes out the other side is a new work rather than a copy of any single song.
That argument is where Australian law gets awkward. Copyright is built around the act of reproduction, so it is well suited to catching someone who uploads a pirated album. It is far less comfortable with a process that copies millions of songs once, during training, and then produces outputs that do not obviously reproduce any of them. An artist who suspects their catalogue was used often cannot prove it, because the training data is a closely guarded secret. Even when they can, the remedies are slow, expensive and uncertain.
Two views on where the line sits
The technology industry’s position is that training is a transformative, largely invisible step, closer to a person listening to thousands of records and being influenced by them than to outright copying. On that reading, forcing developers to license every track would make competitive model building in Australia commercially impossible, and would simply push the work offshore to jurisdictions with looser rules. Some in the sector point to the promise of AI tools that help musicians demo ideas faster, clear administrative drudgery and reach audiences, and warn against regulation that treats every model as a threat.
The creative sector sees it very differently. Collecting society APRA AMCOS, which represents songwriters and publishers, has consistently argued that training on copyrighted work is a commercial use that should be licensed and paid for like any other. Its chief executive, Dean Ormston, has been vocal that consent, transparency and remuneration are non-negotiable principles, not optional extras. The fear across the industry is not just lost licensing income but substitution: if a streaming platform or an advertiser can generate cheap, serviceable music that leans on the style of established Australian acts, the human artists who created that style get nothing while their market shrinks.
The policy fight behind the headlines
This is not an abstract debate. It has become one of the hottest questions in Australian copyright policy, driven in large part by the Productivity Commission’s review of data and digital technology, which floated the idea of a text and data mining exception that would make it easier to use copyrighted material to train AI. The creative industries reacted with alarm, reading the proposal as a plan to legalise the very scraping they are fighting. Music, screen, publishing and visual arts bodies lined up to warn that an exception would hand the value of Australian culture to largely foreign technology companies.
The Albanese government has so far resisted committing to a broad exception, and the Attorney-General’s Department has been running its own consultation on AI and copyright. The likely landing point, if there is one, is a licensing framework or a transparency regime that forces developers to disclose what they train on, rather than a blanket free pass. Nothing has been legislated, which is precisely why artists feel exposed right now.
What it means for Australia
The stakes here are sharper than they might be elsewhere, because music is one of the few cultural exports where a mid-sized country punches well above its weight. Australian songwriters and performers compete globally, and the royalty streams that flow back home support a long tail of studios, session players, managers and small labels. If training goes unlicensed, the economic logic of that ecosystem starts to erode from underneath.
There is also a sovereignty dimension that lands close to the debates FluentSea has covered around local models and data centres. Most of the large audio models being trained today are built offshore. If Australian recordings become raw material for those systems without payment, the country is effectively exporting its cultural capital for free while importing the finished product back as a service. That is a poor trade for a nation that says it wants a genuine domestic AI industry rather than a permanent role as a data supplier.
For working musicians, the practical advice in the meantime is unglamorous: register works properly, keep records, watch the metadata attached to releases, and lean on collecting societies to press the case collectively, since individual litigation against opaque model builders is beyond almost everyone’s means.
What’s next
Expect the pressure to build on two fronts. Legally, the outcome of the government’s copyright consultation will decide whether Australia tilts towards a mining exception, a licensing mandate or a transparency rule, and each option produces very different winners. Commercially, the bigger record companies are already striking licensing deals with AI developers overseas, which could create a template that leaves independent Australian artists to negotiate from a much weaker position. Either way, the comfortable assumption that copyright would protect the likes of Kylie and Barnesy by default has not survived contact with generative AI, and closing that gap is now squarely a job for policymakers rather than the courts alone.
Sources: Startup Daily.
