Ask an Australian songwriter how they feel about artificial intelligence and the answer tends to arrive quickly and bluntly. Musicians have been among the loudest critics of generative models that hoover up recorded catalogues to learn how to spit out new tracks on demand. Yet the anger, however justified it feels to the people making the music, runs headlong into an uncomfortable reality: the law that is supposed to protect a song offers musicians far less cover than it offers a novelist or a journalist, and closing that gap is turning out to be genuinely difficult.
That is the tension at the heart of a recent feature published by ArtsHub, which examined why performers and composers are likely to find it much harder than authors to prove that an infringement has even taken place. The problem is not that musicians care less or that the technology treats them more gently. It is that the mechanics of how music is made, and how copyright applies to it, make the harm slippery to pin down in a courtroom.
Why a song is harder to defend than a book
When a large language model reproduces a chunk of a book close to verbatim, the evidence tends to speak for itself. A reader can lay the passage from the training text alongside the machine’s output and see the overlap. Litigation launched overseas by authors and news publishers has leaned heavily on exactly that kind of side-by-side comparison, and it has given writers a reasonably clear line of attack.
Music does not behave that way. A song is a layered thing, built from a melody, a chord progression, a rhythm, a production style and a particular vocal timbre, and copyright in Australia protects some of those elements while leaving others in a grey zone. Chord sequences and common progressions have long been treated as the shared furniture of popular music, which is why so many hits sit on the same four chords without anyone being sued. When an AI model absorbs thousands of tracks and then generates something that sounds like a given artist without lifting any single identifiable phrase, the musician is left trying to prove a resemblance rather than a copy, and resemblance is notoriously hard to litigate.
There is also the question of what the machine actually did. Training a model on a recording is not the same as releasing that recording, and the legal status of the copying that happens during training remains unsettled in Australia. If a system ingests a catalogue, learns its statistical patterns and then produces output that competes with the original artist in the market, the harm can be real and commercial even when no listener could point to a stolen bar of music. Existing copyright doctrine was simply not designed for that scenario.
Two ways of seeing the same problem
For the artists and their representative bodies, the argument is about consent and payment. Songwriters and performers want the right to say no to having their life’s work used as raw material, and where they cannot say no, they want to be paid. Collecting societies and industry groups have pushed for a licensing model in which AI developers must seek permission and compensate rights holders before training on Australian repertoire, in much the same way radio play and streaming already generate royalties. From that vantage point, the current situation looks like a straightforward transfer of value from creators to technology companies, dressed up as innovation.
The technology side sees it differently. Developers argue that training a model on publicly available material is closer to studying and learning than to copying, and that pattern recognition across millions of works produces something genuinely new rather than a reproduction of any one input. They warn that an overly strict licensing regime could make it impossible to build competitive models in a small market like Australia, pushing development offshore to jurisdictions with looser rules and leaving local companies behind. Between those two positions sits a policy question that no Australian government has yet resolved: whether the country should carve out a text-and-data-mining exception of the kind being debated in Britain and the European Union, or hold the line and force AI firms to license what they use.
The Australian stakes
This is not an abstract overseas fight that happens to interest local musicians. Australia has a deep and export-oriented music sector, and organisations such as APRA AMCOS collect and distribute royalties on behalf of tens of thousands of songwriters and publishers. If AI-generated music trained on Australian catalogues begins to flood streaming platforms, the downward pressure on an already thin income base could be severe, particularly for the mid-tier and emerging artists who never had major-label leverage to begin with. A working musician who earns a modest living from sync licensing, live performance and streaming has a great deal to lose from a technology that can approximate their sound for a fraction of the cost.
The policy backdrop matters too. Canberra has been circling the intersection of copyright and AI for some time, with the Attorney-General’s Department running consultation on how the framework should adapt and the Productivity Commission floating the idea of exceptions that would make it easier to train models on copyrighted material. Creative sector groups have pushed back hard against anything resembling a blanket carve-out, arguing that it would legalise the very conduct artists are objecting to. How that debate lands will shape whether Australian musicians end up with enforceable rights or merely a grievance.
What happens next
The near-term battle will be fought on two fronts. In the courts, expect test cases that try to stretch existing copyright concepts to cover AI training and output, though the difficulty of proving musical infringement means litigation is a blunt instrument here. In policy, the more consequential fight is over transparency: forcing AI developers to disclose what they trained on would at least let artists know whether their work was used, which is currently close to impossible to establish. Without that visibility, the whole question of infringement stays theoretical.
There is also a market response taking shape, with some artists and labels exploring licensing deals that put a price on their catalogues rather than waiting for the law to catch up. That pragmatism may prove more effective than litigation in the short run, but it also risks entrenching the idea that training data is simply there for the taking so long as someone eventually signs a cheque. For now, Australian musicians are left in an awkward spot: convinced they are being wronged, and largely unable to prove it. Closing that gap will take clearer law, real transparency obligations, and a political willingness to decide whose interests come first.
Sources: ArtsHub
