No Exceptions, Please: OpenAI and Anthropic Lobby Australia to Soften Its AI Copyright Ban

In the global fight over what AI models are allowed to learn from, a new front has opened in Canberra. OpenAI and Anthropic — two companies more famous for racing each other to the frontier than for filing joint-adjacent paperwork — have both lodged submissions with an Australian parliamentary inquiry, asking the government to reconsider its refusal to let AI developers train models on the country’s creative content. Reuters, which saw the documents, reported the filings on 22 September 2026.

The backdrop is unusually clean, at least by copyright-war standards. In October 2025, the Australian government stated plainly that it would not introduce a text-and-data-mining (TDM) exception into its copyright legislation. No carve-out, no fair-use-style escape hatch: if you want to train a commercial model on an Australian novel, song, photograph or newspaper archive, you need permission from the rights holder first. For foreign labs headquartered in San Francisco, that effectively adds up to a prohibition on training within the country.

Two labs, two very different pitches

What makes the story interesting is not that the labs are lobbying — that is Tuesday — but how differently they are lobbying, and how explicitly both have tied their policy asks to money already committed on Australian soil.

Anthropic’s filing is the more concrete trade. The company does not contest the politics: it acknowledges in its submission that broad copyright exceptions failed to generate support and have been ruled out by ministers. What it proposes instead is a “narrow form of conditional approval” — a pathway that would let training proceed under defined circumstances, with obligations attached. Anthropic indicated it would be open to conditions such as investment in Australian creators or other support for the country’s cultural work. Permission, in other words, but not for free.

OpenAI’s pitch is framed more broadly. It calls for a “balanced copyright framework” that would let models learn from publicly available information, while offering rights holders opportunities to collaborate with AI companies — language that stops short of Anthropic’s conditional-approval framing but lands on the same destination from a different angle: some legal route to training without a licence negotiation for every work.

Both companies connected the question to their infrastructure plans in Australia. OpenAI holds an offtake arrangement with data-centre developer NextDC for a Sydney facility, and Anthropic was named just last week as a partner on a Queensland project. The not-so-subtle message: copyright settings and investment decisions are linked.

The FOI paper trail

Anthropic has form here, and the paper trail is now public. Briefing notes released under Australia’s Freedom of Information laws show that chief executive Dario Amodei personally raised the copyright issue with Treasurer Jim Chalmers in April. Officials noted at the time that Anthropic would argue its investment in Australia depended on clarity about copyright settings. The parliamentary submission is the same argument, moved from a private meeting into the formal record.

There is also a courtroom shadow over the whole debate. In the United States, Anthropic agreed in 2025 to pay $1.5 billion to settle a class action brought by authors over pirated books used in its training data — a settlement that followed a judge’s ruling that training on lawfully acquired books could qualify as fair use. The company has therefore now argued, in different jurisdictions, both that training can be legal without permission and that it will pay billions when it wasn’t. Australia’s creators have noticed.

Creators push back

The creative industries are not sitting this one out. In July, Greens senator Sarah Hanson-Young — who chairs a separate Senate inquiry into AI and data centres — said AI companies wishing to train on copyrighted material should expect to pay for it. Authors, publishers and creative bodies have consistently opposed exemptions that would let commercial systems absorb their work without consent or compensation, and the government now finds itself arbitrating between those concerns and the investment case for AI infrastructure.

The stakes are not small. Economists quoted by Reuters put the value of the AI sector at hundreds of billions of dollars globally by 2030, and governments everywhere are competing for a share. Australia wants the data centres, the jobs and the strategic relevance; the question is what it will trade for them.

What happens next

The submissions now sit with the Joint Select Committee on Artificial Intelligence, established by both houses of parliament in August, which is due to report in November 2026. Its recommendations will signal whether lawmakers see room for the narrower training permissions the labs want — after the government has already rejected the broad version. In parallel, the Attorney-General’s Department has said it is not considering a TDM exception at all; its Copyright and Artificial Intelligence Reference Group is instead examining fair licensing, certainty for AI-generated material, and lower-cost enforcement. ABC reported on 16 September that the government has been consulting creators, media organisations and AI companies about possible control-and-payment models — options that leaked material suggests drew criticism from creators and political opponents alike. Comprehensive AI and data-centre regulation is expected to be introduced next year.

Why this matters beyond Australia

Three reasons this filing deserves attention even if you never touch Australian law.

First, it is a test of the “conditional exception” model. The great copyright standoff of the AI era has mostly been binary: unrestricted training (the labs’ old position) versus consent-and-payment (the creators’ position). Anthropic’s conditional-approval language — training allowed, but tied to investment in local creators — is an attempt to build a third lane. If Canberra takes it seriously, expect the template to travel to other small-to-medium markets being courted for data-centre investment.

Second, it shows infrastructure as leverage. Both labs explicitly linked training rights to data-centre plans. As AI investment becomes a geopolitical bargaining chip, copyright policy is increasingly set in the same room as energy deals and land grants. That is a structural change in how these debates get decided.

Third, the compliance baseline hasn’t moved. As legal analysts have been quick to note, these are inquiry submissions, not enacted rules. No new exemption exists. Publicly available content still does not automatically permit training. For any organisation building with AI in Australia — or training on Australian material anywhere — the operative task is documenting the rights basis for every dataset: source, licence, territory, permitted purposes, and a removal path. Policy uncertainty is not a permission slip.

Canberra has said no to the big ask. The labs are now negotiating the size of the small one. November’s committee report will tell us how small.