AI copyright reform: Farquhar’s Australian training-data fight

Atlassian co-founder Scott Farquhar has pushed Australia’s copyright debate into a more practical corner: if local AI builders cannot train models on domestic content without negotiating thousands of licences first, can Australia credibly claim it wants a sovereign AI industry?

At the Australian Financial Review’s AI Summit, Farquhar argued that current settings make local model training close to impossible. Legal friction was only part of the case. His larger point was about investment, data centres, start-ups and whether Australian companies build AI systems here or buy them from offshore platforms.

Creators heard a different argument. APRA AMCOS and music-industry executives read the proposal as a bid to lower the price of training data, not as a necessary fix for broken law. By the third day of the dispute, “AI copyright reform” had become shorthand for something more concrete: who gets paid when creative work becomes infrastructure.

“If I train in Australia, I need to cut a deal with every single recording artist in the entire world … it is impossible.”
Scott Farquhar, speaking at the Australian Financial Review AI Summit

That line, carried in Farquhar’s AFR remarks, fits the Tech Council’s broader claim that AI can lift national productivity only if Australia removes policy friction around talent, investment and deployment. Copyright is now being folded into the same competitiveness file.

Copyright becomes an AI infrastructure question

The strongest version of Farquhar’s argument is not that creators should go unpaid. It is that model training does not look like a normal music, publishing or screen licence when it happens at internet scale. A domestic AI company trying to build a frontier or sector-specific model may need text, code, music, images, video and other data from many jurisdictions before it even reaches a product.

Consider the timing. Australia is already trying to attract capital for data centres, cloud regions and model access. One recent report said AI demand could push Australian power prices higher by 2035, while other commentators have argued the country lacks enough public data on the water and energy impacts of the data-centre rush. Copyright is not the only constraint. It may not even be the biggest one.

Still, copyright is unusually symbolic. Power, land and chips are physical bottlenecks. Copyright is a political bottleneck. It asks whether a country with a deep creative sector but a comparatively small technology base should make it easier for AI firms to copy training material first and settle compensation later.

Rights holders reject that premise. APRA AMCOS chief executive Dean Ormston said AI investment decisions are driven mainly by capital, energy, water and land, not copyright law. Weakening copyright, on that view, would not solve the hardest infrastructure problems. It would just shift value away from artists.

“AI investment decisions are primarily based on access to capital, energy, water and physical land, not copyright law.”
Dean Ormston, APRA AMCOS chief executive

Rights holders see a bargaining fight

For the music industry, Farquhar’s “impossible” claim attacks the premise of collective licensing. The whole point of a collecting society is to turn millions of individual works into commercial licences. Slow, expensive, uneven. But not imaginary.

Annabelle Herd, the chief executive of ARIA, pushed back on the idea that an AI company would need to strike deals with every recording artist individually. In comments carried by The Music Network, she said major licensing routes already cover much of the global recording catalogue.

“You don’t have to go into every single territory … I could give you five phone numbers, and you would have 80% … of the world’s sound recordings licenced.”
Annabelle Herd, ARIA chief executive

That answer does not settle the whole AI problem. Sound recordings are not books, journalism, software repositories or public images. A licensing path that works for music may not scale cleanly across every category of training material. Nor does it say much about small creators, independent publishers or works where ownership is contested.

Even so, the rebuttal changes the frame. If licensing is merely hard, the policy answer may be better collective licensing, clearer transparency rules and standardised negotiations. If licensing is genuinely impossible, the answer may be a broader exception, compulsory licence or statutory fund. Those are different choices.

Mediaweek’s reporting on the backlash showed why creators are wary of a one-off settlement model. Rights bodies do not want AI platforms to pay once into a pool and then treat Australian work as training commons. They want continuing licensing, auditability and bargaining power as models improve.

Canberra has to pick the mechanism, not the slogan

Canberra’s task is narrower than the public rhetoric. It need not pass judgment on AI or treat copyright as untouchable. It needs to decide what mechanism lets Australian AI development proceed without telling creators that their work has become free industrial input.

Several routes are available. A narrow text-and-data-mining exception would give AI developers more certainty, but it would be politically difficult after rights holders mobilised against similar carve-outs overseas. A compulsory licence could allow training while setting payment rules, though it would require a valuation method for works that contribute unevenly to model performance. Transparency rules could force AI firms to disclose what they trained on, but disclosure alone does not create a licence.

Leaving the law largely as it is would be a choice too. Major AI firms would keep negotiating from offshore. That may protect existing rights on paper. It may also leave Australian start-ups with less leverage than global platforms that can afford lawyers, compute and private licensing teams.

Farquhar’s broader policy campaign makes that risk explicit. In a separate Australian Financial Review summit discussion, he argued that capital-gains tax changes had doubled the effective tax rate faced by start-up founders. Copyright sits beside that complaint. Both arguments say Australia is making it harder than it needs to be for technology companies to form, scale and stay onshore.

That does not make the copyright argument correct. It does explain why the Tech Council is pressing it now.

The local answer may be less dramatic than either side wants

Australia is unlikely to win the global AI race by copying the largest US platforms. It does not have their capital base, compute scale or domestic market. Its stronger case is narrower: regulated industry tools, public-sector AI, health and education systems, mining, agriculture, legal services and software products built around Australian data and compliance needs.

Those use cases make copyright more important, not less. A local model trained for Australian law, media, education or culture will need Australian material. If that material is locked behind uncertain rights, local firms face a compliance burden. If it is opened too broadly, creators face a revenue threat from companies that may later sell model access back to the same market.

A workable settlement will probably look less dramatic than either side wants. Standard terms. Collective bargaining where possible. Opt-out or opt-in rules that are actually enforceable. Transparency duties that do not expose trade secrets by default. Such a structure would disappoint AI firms that want clean immunity, and rights holders who want every use negotiated case by case.

That is usually where durable policy lands.

Farquhar has opened a fight about whether Canberra can make training data legible enough for investment without making it so cheap that creators become the only part of the AI supply chain expected to subsidise everyone else.

For Australian AI builders, the answer will shape more than copyright compliance. It will decide whether “build here” means training on Australian work under Australian rules, or assembling foreign models around the edges and calling that sovereignty.