Elton John Joins Australian Copyright Campaign as Proposed AI Training Licences Spark Fight in A$70.3bn Sector
Elton John's ARIA appeal meets reported collective licensing options and an expected AI legislative package. The commercial fight turns on whether training payments can follow models across borders.
Elton John appeared at Sydney's ARIA Awards on 18 November 2026 by prerecorded message, demanding permission for artists. "Don't let AI use any Australian artist's work without full permission." He backed Not AI's To Take, which wants permission and agreed payment. We would resist pricing this campaign as a universal royalty on foundation models. The immediate trade is over where model developers train and which rights holders collect. That matters more than pricing every book a model learns from.
Australia can impose a payment duty for training at home while developers can conduct the training abroad. A Movable Mandate is a copyright payment duty that alters model costs only when regulated training cannot leave the jurisdiction or the payment follows models into the destination market. Copyright Agency chief executive Josephine Johnston told Parliament on 6 October 2026: "To date, the training has all happened overseas." Location determines whether a domestic charge can be collected. Our baseline concerns location before an industrywide royalty.
Attorney-General Michelle Rowland ruled out a text and data mining exception on 26 October 2025 and reaffirmed it on 26 March 2026. The Copyright and AI Reference Group is weighing a paid collective licence against continued voluntary deals. A broader AI legislative package is expected in late 2026 or early 2027, with copyright disputed; no new framework has been enacted. One puts an administrator between creators and developers; the other leaves owners negotiating. Both sell legal certainty, and both redistribute negotiating power. Yet a payment framework aimed at training in Australia meets a business able to do its training elsewhere. An entitlement and a collected fee remain different things.
Guardian Australia reported that leaked options tabled in the Senate on 15 September 2026 contemplated broad deals with rights holder organisations or payments through a new central body. The body would distribute money to "registered" rights holders. Societies identifying works and processing claims would have access to payments; unregistered creators would face another hurdle. The actor selling developers clean rights across entire categories of works would gain leverage. A proposed licensing scheme can end up pricing administration as much as creation.
Not AI's To Take says Australia's creative and media sector contributes A$70.3 billion and employs almost 600,000 people. Those figures describe the sector, rather than a royalty pool. The tempting price comes from Anthropic's US$1.5 billion copyright settlement covering a reported 482,000 books. In a settlement scenario, US$1,500,000,000 over 482,000 books is US$3,112 gross per book; the settlement site says approximately US$3,000 per work before adjustments for interest, approved costs and fees. A firm that treats this as a lawful training tariff mistakes a piracy settlement for a recurring input cost.
Judge William Alsup drew that boundary on 23 June 2025. He held that using the books to train Claude "was exceedingly transformative and was a fair use" under US law. Anthropic had also downloaded over seven million pirated copies for its library, conduct that received separate treatment. Training incurred zero royalty liability under that ruling; the piracy claims produced the US$1.5 billion settlement. Those activities have different prices, even when the same books are involved.
Anthropic offers an economic alternative. Reuters reported its request that "the government could permit AI model training through a narrow form of conditional approval", with opt-out signals in robots.txt. OpenAI sought "a balanced copyright framework that allows models to learn from publicly available information". Developers could train broadly without individual licences while owners retain a reservation mechanism. At the 2025 US court's horizon, training royalties were zero. For a company placing its next training run, a proposed Australian mandatory charge turns offshore compute into the cheaper legal route, at least for the training charge itself.
Independent senator David Pocock urged delay, citing more than 100 US copyright and AI training cases. European lawmakers supplied an actual alternative in Article 4 of their 2019 copyright directive. It creates a text and data mining exception unless rights holders expressly reserve their works, including through machine-readable means. That arrangement enables training without per-work clearance fees while allowing opt-outs. It is the outcome Anthropic seeks; Australia's government has already ruled out the comparable exception.
Johnston's October 2026 testimony challenges our geography argument. She described an emerging global licensing market that can benefit Australian creators represented by global publishers, despite training overseas. Rights can travel through contracts even when domestic copyright duties stop at borders. APRA AMCOS executive Jonathan Carter made identifiable training datasets central to a workable licensing system. Global publishers could charge for foreign training without relocating model runs, using rights negotiated across more than one national market. That rewards firms assembling verifiable rights beyond Australian data centres.
Still, a worldwide contract is not an Australian statutory levy. Johnston's 6 October 2026 evidence was that training has all happened overseas. For that offshore training an onshore mandate would reach nothing, although global contracts could still reach offshore activity. The distinction matters because Australia's rejected exception cannot be replaced by a US court's fair-use conclusion. To change that arithmetic for offshore runs, a bill would need to attach payment to domestic model supply, rather than Australian training alone. Only then would the cost follow the developer.
Consider the infrastructure offers. OpenAI has an offtake agreement with NEXTDC for a planned Sydney data centre, and Anthropic agreed to use part of a proposed A$32 billion Queensland hub built by Zerra DC. A$32 billion describes the facility, not Anthropic's investment. The ABC understands Anthropic intends to use the site from 2027 to answer Claude users' questions, rather than train; the lease requires foreign investment approval. Local inference would serve customers while leaving training elsewhere. Training and supplying models would create different Australian payment liabilities.
Australia's News Media Bargaining Code is revealing. It took effect in March 2021; a December 2022 Treasury review counted more than 30 deals compensating media outlets. Google and Meta needed Australian users and news, making the domestic market the bargaining lever. Training compute can move, so an Australian charge tied solely to a local training run has a weaker grip. A payment duty attached to supplying foreign-trained models into Australia would resemble the Code's leverage. The Movable Mandate would become a market access cost instead of a relocation incentive.
We would price the proposals as a fight over training locations, rights administration and market access, rather than a worldwide copyright tariff. We would be wrong if, by 30 June 2027, the Australian Parliament bills list recorded an introduced bill that would create an enforceable AI training payment duty covering offshore training or models trained abroad and supplied into Australia.