Canberra just moved the goalposts on AI

Canberra has shifted the rules for anyone using artificial intelligence (AI) in Australia. If you run a startup, a marketing team, or a product built on a United States model, you now have a compliance problem you did not have a week ago.

On 15 July 2026, Prime Minister Anthony Albanese used a speech at the University of Sydney to declare that unlicensed training of AI models on Australian books, music, art or news amounts to "theft." The same day, the government stood up a new Office of AI inside the Department of the Prime Minister and Cabinet, and committed to introducing AI legislation in early 2027. The proposed mandatory standards go to National Cabinet in August 2026 for agreement with the states.

The move locks in a direction Attorney-General Michelle Rowland set in October 2025, when she formally ruled out a text-and-data-mining exception to the Copyright Act. That is the kind of carve-out US and European Union developers rely on. 

Australia's existing "fair dealing" provisions are narrower than US fair use and, as both MinterEllison and White & Case have noted, do not stretch to commercial AI training at scale. Put plainly, any model trained on Australian copyrighted material without a licence is, on the current reading, already infringing.

This matters for downstream users, not just model builders. Australian businesses running Claude, ChatGPT, or Gemini through application programming interfaces (APIs) are increasingly being asked by their lawyers to review indemnity clauses in vendor contracts, because liability for training-data infringement now flows through the stack.

Anthropic,, has been vocal in the debate. According to a Treasury briefing note released under freedom of information laws and reported by the Australian Financial Review, the company has tied a proposed A$21.6 billion (US$15 billion) Australian data-centre investment to "clarity of copyright settings." Anthropic settled a US class action in 2025 for US$1.5 billion over its use of pirated books in training, the largest copyright settlement in US history. That settlement is the practical benchmark local courts will look at if Australian rights-holders sue.

For Australian startups the picture cuts both ways. Compliance costs are about to rise through vendor due diligence, contract renegotiation, and potentially licensing fees for any locally sourced training data. But clear rules can also work as a moat. 

Institutional capital, particularly the super funds now writing venture cheques, dislikes legal ambiguity. A compliant Australian AI stack, built on properly licensed data, may attract exactly the kind of risk-averse capital that has stayed on the sidelines.

The date to watch is the National Cabinet meeting in August 2026. What the states agree there sets the shape of the data-centre rules, and signals how hard Canberra intends to hold the copyright line when legislation is drafted for 2027.

If you use AI tools in your business, ask two questions this week. First, does your contract with your AI vendor indemnify you for training-data infringement claims? Most standard terms do not. Second, if you build product features on top of a foundation model, do you have visibility on what that model was trained on? If the answer is no, you are carrying risk you have not priced. The regulatory direction is set, and it will not soften. The businesses that do the compliance work now, before the legislation lands in 2027, will be the ones raising money while their competitors are renegotiating contracts.