Australia is the AI copyright outlier, and the world's music publishers are warning it

The global body for music publishers has a blunt message for Canberra: everyone else said no. The International Confederation of Music Publishers (ICMP), which represents Sony, Warner and Universal's publishing arms alongside thousands of independents, told Guardian Australia on 19 October that 62 countries have considered changing copyright law for AI training. By its count, none of the other 61 has introduced a commercial text and data mining (TDM) exception, and Australia appears to be the only one still entertaining one. That count is ICMP's own tracking and we couldn't check it country by country, but it lines up with what the larger markets have publicly done.

What is ICMP actually claiming?

Three things. First, that AI companies are pressuring Australian policymakers with an investment threat: change the law or we won't build here. ICMP managing director John Phelan described the goal as asking lawmakers to make theft legal ("legitimise larceny", in his words to the Guardian; The AI Musicpreneur also covered the interview).

Second, that Australian music is already inside these models. ICMP showed the Guardian that Google's Gemini could reproduce lyrics by Kylie Minogue, G-Flip and Nick Cave. The Guardian notes that a key industry claim is that models are not trained on Australian content because of local copyright rules.

Third, that the training pipelines are mundane. Phelan pointed to a dataset reported by The Atlantic in June. The source is LAION-DISCO-12M, published by the non-profit LAION in November 2024. It lists metadata for 12,648,485 songs (title, artist, album) paired with YouTube links. LAION says it doesn't distribute the audio itself and advises against industrial use. The Guardian reports it includes tracks by Minogue, Cave and Hilltop Hoods.

Why is Australia back in play?

Because a parliamentary committee reopened a question the government had closed. In October 2025, Attorney-General Michelle Rowland said the government would not consider a TDM exception and pointed the sector towards licensing talks through its Copyright and AI Reference Group. In July 2026 Prime Minister Anthony Albanese said it would be "theft" if artists lost control over whether their work trains AI and what they're paid for it, per the Guardian.

Then, in a joint letter this month, the chair of the Joint Select Committee on Artificial Intelligence, Labor MP Jo Briskey, and her deputy, Liberal senator Jonathon Duniam, wrote to Albanese recommending the government "examine current copyright settings" while keeping creators in control and paid. They called the situation a stalemate. The argument they relayed is that US models already train on copyrighted work without paying, while companies hold back frontier training in Australia until the law changes. Independent senator David Pocock has raised concerns that the committee's position was predetermined after staffer edits were seen in the letter. Briskey rejects that.

Anthropic's special envoy Jeffrey Bleich gave the committee the industry case: you can't license the whole internet, and under current Australian law training could be found illegal unless a licence were obtained from every rights holder. OpenAI's Adam Cohen argued US fair use protects new technology, while citing the deals OpenAI has struck with News Corp and the Guardian.

The committee was set up on 20 August 2026 with copyright explicitly in scope. Submissions closed on 14 September and the final report is due by 30 November, according to Clayton Utz's summary of its terms.

What the UK's reversal shows

The UK is the closest precedent, and it ran the other way. In 2024 its preferred option was a broad data mining exception with an opt-out for rights holders. Its report to Parliament of 18 March 2026 records 11,520 consultation responses. Only 3% backed the opt-out exception, while 81% preferred strengthening copyright so that licensing is required in all cases. The government concluded that the opt-out exception "is no longer the government's preferred way forward".

Read the next line too, though. The UK didn't adopt the licensing-required option either. It promised more evidence-gathering, work on transparency and labelling best practice, and to monitor rather than intervene in the licensing market. So the artist campaign stopped a bad law, but it didn't win a good one.

What it means for artists

For Australian musicians the stakes are concrete:

• Consent. An exception would turn "ask first" into "opt out if you can", or remove the choice entirely. Albanese's own framing, control plus payment, is the test any committee recommendation should be held to.
• Income. The reference group has debated a paid collective licensing scheme against voluntary deals, per The Lawyer Magazine. An exception would undercut that negotiation before it concludes, because there's nothing to license if training is free.
• Lyrics are songwriter income. The Gemini evidence is about lyrics, which are the publishing side of a song. That is why a publishers' body, not a label group, is leading this fight, and why songwriters who don't perform should pay attention.

The response online has mostly been alarm at the names involved. Cybernews, for example, amplified the Minogue and Cave training claim on X. Name recognition helps, but the people most exposed are the thousands of working Australian writers with catalogues on YouTube and no legal team.

What we think

The investment threat is the weakest part of the AI companies' case, and the Gemini lyrics show why. If the models already reproduce Australian songs, the "we won't train here" warning isn't protecting Australian creators from anything. It is asking Australia to retroactively bless copying that has already happened abroad. And as ARIA's Annabelle Herd told the committee, an Australian carve-out wouldn't stop at Australian songs. A lone exception in one mid-sized market would make it a legal training haven for the world's catalogue. That's why a global publishers' body cares this much about one country's committee report.

The second-order effect matters more than the headline. If Australia carves out an exception, every licensing negotiation elsewhere gets harder, because AI firms gain a fallback jurisdiction. If it holds, the 61-of-62 pattern becomes close to a global norm, and the bargaining moves to price and transparency, which is where artists need it.

The UK is also a warning. Winning the vote on exceptions isn't the same as getting paid. If Australia's committee ends at "no exception, keep talking", artists are left where they are now: their work already in the datasets, and no money. The better ask before 30 November is a dated licensing framework with training-data transparency attached. Submissions have closed, so the route now is MPs, committee members and peak bodies. That leaves weeks, not months.