The AI in Australia's Interests Speech: What It Means for Copyright, Data Centres and AI Training

A Speech Billed as a Turning Point for Australian AI Policy

On 15 July 2026, Prime Minister Anthony Albanese delivered an address titled “AI in Australia’s Interests” at the University of Sydney, describing it as a defining moment for how Australia approaches artificial intelligence.¹ The speech announced the immediate establishment of a new Office of AI within the Department of the Prime Minister and Cabinet, a commitment to develop a set of Australian Standards for AI, and a firm statement that Australian creators must retain ownership and control of their work.¹²

The announcement was widely covered as a significant, world-first attempt to bring data centre regulation, copyright protection and national security considerations into a single coordinated framework.³⁴ For Queensland businesses, creators and technology developers, the practical legal position is more complicated than the headline suggests, and several important questions remain unresolved.

The logical conclusion of the current legal framework is stark: until the promised legislation is enacted, training a large language model or other frontier AI system in Australia using internet-scale data is effectively commercially infeasible. There is no text and data mining exception, no government-sanctioned licensing scheme, and no judicial authority confirming a fair dealing defence would succeed. Any entity that proceeds faces an unacceptable and essentially unquantifiable copyright liability exposure — one that could attach to every work in the training dataset. The government has signalled its intent clearly, but intent is not law. Businesses and developers should treat the period between now and early 2027 not as a green light but as a legal gap during which proceeding with large-scale AI training in Australia carries significant unmitigated risk.

What Was Actually Announced

The Office of AI

Effective from 15 July 2026, the Prime Minister established the Office of AI within his own Department of the Prime Minister and Cabinet.¹⁵ According to the government’s own description, the office is intended to coordinate the design of new Australian Standards for AI and to bring together AI-related work being undertaken by ministers across government, including energy, copyright, productivity, employment, education, online safety, defence and national security.⁵⁶ The office works closely with the Minister for Industry and Innovation, Tim Ayres, and the Assistant Minister for Science, Technology and the Digital Economy, Andrew Charlton.¹⁶

Australian Standards for AI and Data Centre Obligations

The speech announced that the government will bring existing expectations for large AI data centres, first outlined in March 2026, into a single, mandatory regulatory framework.¹ Key elements include:

• A legal obligation for large-scale data centres to underwrite new power supply and pay their full share of grid connection costs, so that costs are not passed on to households or businesses.¹⁷
• A requirement for data centres to be net contributors of energy to the grid, rather than net users, and to build new renewable generation and firming capacity.¹
• Mandatory requirements to minimise water use and pay for any additional water infrastructure required, reflecting Australia’s status as the driest inhabited continent.¹

The Prime Minister indicated he would seek agreement from Premiers and Chief Ministers on this approach at National Cabinet in August 2026, with legislation targeted for introduction to Parliament in early 2027.¹⁸

Copyright Protection for Creators

The speech contained the government’s clearest public statement yet on AI and copyright. The Prime Minister stated that “Australian writers, musicians, artists and journalists must retain ownership and control of their work,” and that “no company should use Australian books, music, art or news to build or train AI without the artist’s control.”¹ This builds directly on the government’s decision in October 2025 to reject a proposed text and data mining exception to the Copyright Act 1968 (Cth), which would have permitted AI developers to train on copyrighted material without a licence, a mechanism used in some form in the European Union and Singapore.⁹¹⁰ Attorney-General Michelle Rowland confirmed at the time that the government would not weaken copyright protections for the benefit of AI development, and that a body of consultation work remained ahead on how the copyright environment would apply to AI specifically.¹¹

How This Fits Within Existing Australian Copyright Law

What Copyright Protects, and What It Does Not

Australian copyright law protects the original expression of ideas, not the underlying ideas, facts, concepts or information themselves.¹² This distinction, sometimes called the idea-expression dichotomy, means that copyright subsists in the specific words of a book, the specific code of a computer program, the specific notation and recording of a musical work, or the specific composition of an image or film, but does not extend to the broader concept, theme or factual content those works may express or convey.¹²¹³ The Copyright Act 1968 (Cth) protects defined categories of subject matter, including literary, dramatic, musical and artistic works, along with sound recordings, films and broadcasts.¹³

This distinction is directly relevant to AI training. Developers of AI systems have argued that training a model involves learning statistical patterns from text and images, comparable in principle to how a person learns style and structure from wide reading, rather than storing and republishing specific protected expression.¹⁴ Rights holders and their representative bodies argue that the training process necessarily involves reproducing copyrighted works during the data collection and processing stages, which is capable of infringing the reproduction right regardless of what the resulting model retains.⁹¹⁴ This is not a novel dispute. Litigation overseas, discussed further below, has already tested both positions without producing a single settled answer.¹⁵

The Text and Data Mining Decision

Australia’s current legal position on AI training was set in October 2025, when the government confirmed it would not introduce a text and data mining exception to the Copyright Act 1968 (Cth), despite this being recommended in the Productivity Commission’s interim report on harnessing data and digital technology.⁹¹⁶ The practical effect of this decision is that, absent a specific licence or an existing exception such as fair dealing for research or study, using copyrighted material to train an AI system in Australia is likely to infringe copyright.¹⁰¹⁷ This has been described by commentators as leaving Australian copyright holders in a comparatively strong legal position relative to jurisdictions with broad text and data mining exceptions, but it has also raised concerns about whether frontier-scale AI development remains commercially viable domestically under these settings.¹⁸

The Evidentiary Problem: How Would Unauthorised Training Actually Be Proven

Why AI Training Does Not Leave a Traditional Paper Trail

A significant practical difficulty with enforcing copyright protection in the AI training context is that trained AI models do not retain the original source material in any directly retrievable form.¹⁹ Training a large language model compresses vast quantities of text into a set of numerical parameters, commonly described as weights, from which the specific source documents used cannot be recovered or inspected directly.¹⁹ This differs fundamentally from traditional copyright infringement scenarios, where a claimant can typically point to a physical or digital copy of the original work.

In practice, this means that establishing whether a specific work was used in training generally requires prompting the model and assessing whether its output closely resembles the work in question, a form of circumstantial inference rather than direct proof of copying.¹⁵¹⁹

Overseas Litigation as a Guide

Australian courts have not yet been required to resolve a dispute of this kind, but overseas litigation offers a useful guide to how such evidentiary questions might be approached. In the United Kingdom, Getty Images brought proceedings against Stability AI, alleging that its image-generating system had been trained on Getty’s copyrighted photographs without permission.¹⁵ A significant portion of that litigation turned on whether the system’s outputs bore sufficient resemblance to Getty’s watermarked images to support an inference that they had formed part of the training dataset.¹⁵

In the United States, authors and publishers brought proceedings against Meta over its use of copyrighted books in AI training. Court records in that litigation showed that Meta had obtained a substantial volume of material from unauthorised sources.²⁰ Despite this evidence of unauthorised acquisition, a federal court found in 2025 that Meta’s use of the material constituted fair use under United States law, principally because the plaintiffs were unable to establish that the use had caused measurable market harm.²⁰ Australia does not have a general fair use defence equivalent to the United States position, relying instead on more narrowly defined fair dealing exceptions, so a directly comparable outcome should not be assumed to follow automatically in an Australian court.¹²²⁰ The broader lesson from both cases is that establishing that training occurred is only the first evidentiary hurdle, and clearing it does not automatically result in a finding of liability or an enforceable remedy.¹⁵²⁰

Public Records Containing Embedded Copyrighted Material

A further practical complication specific to Australia concerns public records such as Hansard and court transcripts. These are generally treated as public and broadly free to use, but frequently contain quoted copyrighted material within them, including song lyrics tendered as evidence, extracts of journalism cited in judgments, or literary passages read into parliamentary debate.²¹ An AI developer collecting such public records on the reasonable assumption that they are safe to use may inadvertently collect copyrighted material embedded within them, with no straightforward technical method currently available to separate a public record’s own content from privately authored material quoted inside it. This issue has not been directly addressed in the government’s announcements to date.

Compliance Pathways Currently Available

In the absence of a text and data mining exception, an Australian entity seeking to train or fine-tune an AI model while managing copyright risk currently has a limited number of realistic options.

Businesses considering any AI training or fine-tuning activity involving Australian copyrighted material should treat this as an area of genuine legal uncertainty, and should seek advice before proceeding, given the absence of a settled compliance pathway.

What the New Data Centre Rules Do Not Address

Infrastructure, Not Model Provenance

The Australian Standards for AI announced in the speech impose detailed obligations on the physical infrastructure of large data centres, including energy generation, grid connection costs and water use.¹⁷ These obligations are directed at buildings and infrastructure located within Australian jurisdiction, where compliance can be verified through conventional regulatory inspection and reporting mechanisms.

Nothing in the speech or the accompanying Office of AI framework extends equivalent obligations to the training provenance of AI models operated within that infrastructure.¹⁵ A model trained overseas, under whatever copyright and safety regime applied in its country of origin, could in principle be deployed within an Australian data centre without triggering any of the copyright-related obligations discussed in the speech, because the training activity itself did not occur within Australian jurisdiction. This creates a potential asymmetry, whereby Australian-based training activity may be held to a stringent domestic copyright standard, while imported models operating on Australian infrastructure face no equivalent scrutiny of their training history.

Implications for Businesses Considering Cross-Border AI Deployment

Businesses structuring AI deployment across jurisdictions should be aware that the location of a data centre and the jurisdiction in which a model was originally trained are legally distinct considerations under the framework announced to date.¹ Clients should not assume that compliance with Australian data centre standards addresses copyright risk arising from the training history of a model being deployed, particularly where that model was developed outside Australia.

Recommendations for Creators and Businesses

For Creative Industry Clients

• Maintain clear, dated records of authorship and first publication for creative works, to support future licensing negotiations or infringement claims once a formal licensing framework is established.¹¹
• Consider including express text and data mining restrictions in contracts, website terms and licensing agreements, and using metadata or copyright notices to clearly indicate that material is not available for AI training purposes.²³
• Monitor the ongoing work of the Copyright and AI Reference Group, which is expected to shape the practical licensing and remuneration mechanisms flagged in the Prime Minister’s speech.¹¹²²

For Businesses Developing or Deploying AI Systems

• Treat any training or fine-tuning activity involving Australian copyrighted material as carrying genuine legal risk pending further legislative clarity, given the absence of a text and data mining exception.⁹¹⁰
• Seek advice before undertaking large-scale training activity, and consider whether curated or licensed datasets are a more defensible starting point than broad web-scraped material.
• Where deploying AI models within Australian infrastructure, obtain clear representations from vendors regarding the training provenance of any model being used, since Australian data centre compliance does not itself address training-related copyright risk.¹

When to Get Legal Advice

You should seek advice before:

• Training, fine-tuning or licensing an AI model using any Australian creative, journalistic or copyrighted material.
• Entering into contracts with AI vendors regarding the provenance or training history of models to be deployed in your business.
• Responding to a claim, demand or inquiry alleging that your organisation’s AI system was trained on copyrighted material without permission.
• Structuring cross-border AI deployment involving models trained overseas but operated within Australian data centre infrastructure.
• Advising on licensing terms, contractual restrictions or metadata protections for creative works in light of ongoing AI training risk.

Bell Senior Lawyers advises Gold Coast and South East Queensland clients on technology law, intellectual property and AI-related legal risk, including engagement with emerging Commonwealth AI policy and copyright reform.

Related Topics

• Legal Risks of AI for Business – Australia, US and UK
• Is Your Business Data Breach Ready? The NDB Scheme Checklist for Queensland Businesses
• Navigating Tech Startups
• Technology Law Practice Area