Training AI in Australia: Copyright risk in a data centre economy

Data centres are no longer neutral infrastructure

For much of the last decade, data centres were treated as location agnostic hosting environments. AI has now changed that. As models are trained, retrained and then actively used by consumers, AI systems engage in sustained and repeated processes that copy, transform and store large volumes of data. The law that governs such activities is generally where is activity happens where the model is hosted, typically on servers in a data centre, and is therefore governed by the law of that location.

A central copyright question for AI developers is where training and retraining actually occur. Copyright infringement, being a tort, is territorially and determined, at least partly, by where the infringing copying take place, rather than where the model is deployed, the developer is incorporated, the system is used, or where outputs are delivered.

In the AI context, relevant uses of works typically involve repeated copying of data into memory and storage on the servers hosting the system, most often within a data centre. This makes infrastructure location legally significant. When those servers are located in Australia, the relevant acts for copyright purposes, are treated as occurring in Australia, and Australian copyright law therefore applies.

Australia’s copyright stance is a conscious policy choice

Globally, there is no unitary position on whether the unauthorised use of copyright material for AI training should be treated as an infringement, a licence or something akin to fair use. Many jurisdictions are currently seeking to resolve these issues in terms of their own laws. In the United States, courts and commentators continue to debate whether AI training is a fair use of copyrighted material.

Australia has taken a more definitive approach. In late 2025, the Federal Government stated that it does not intend to introduce a broad text and data mining exception for AI training. Instead, attention has been directed towards potential licensing frameworks, improving certainty, and ensuring creators are appropriately compensated¹. This suggests that unlimited uses of copyright material for commercial AI training is unlikely to be permitted or given any statutory form, at least at present.

This clarity is desirable in potentially allowing AI developers to identify and manage copyright risks in Australia early and proactively, rather than leaving matters to be resolved through litigation or regulatory intervention.

Australian copyright law

Under Australian law, copyright includes the exclusive right to reproduce a work in material form². Copyright is infringed where a person, without the permission of the copyright owner, “does” in Australia an act that falls within the scope of the copyright³.

The definition of material form is broad, and includes any form of storage of the work, or a substantial part of it⁴. This includes electronic storage in any form, whether or not the stored material is visible and whether or not it can be reproduced. This is significant in the AI context because model training typically involves a series of technical steps rather than a single operation, often involving multiple acts of copying and storage.

During AI training, copyright issues may arise from acts, including such as:

• copying works into a training dataset;
• reproducing works in memory or storage;
• making temporary copies during ingestion or preprocessing;
• using copyright material for text and data mining.

The use of copyright material in this way is also likely to trigger moral rights issues, such as a content creator’s rights of attribution and integrity and right against false attribution.

Australian copyright law does contain some technical process exceptions. These include certain temporary reproductions made as part of communication processes or reproductions made incidentally as a necessary part of a technical process of using a copy⁵. However, these provisions are narrow and do not operate as general permissions that allow AI to ingest and train on large volumes of third party material for commercial purposes.

Why fair use arguments do not translate cleanly to Australia

Understandably, many AI developers are currently approaching training risk through a United States lens, given the scale of AI development occurring there. That lens is shaped by fair use (including transformative use) doctrine.

By contrast, Australia has no broad fair use defence, and its “fair dealing” provisions are limited to specific purposes. As a result, arguments under US law that AI training “transforms” inputs or produces only statistical outputs do not translate neatly to Australian law. Even if there are no visible outputs from the AI, this doesn’t mean copying isn’t happening during training.

Accordingly, adopting US‑style (or any other jurisdiction’s) reasoning can obscure where legal risk actually arises. For organisations training AI systems in Australia, copyright risks should be assessed by reference to the underlying copying acts, licences and system architecture at the point those acts occur, and in accordance with the local applicable law.

Structuring AI training for the Australian environment

For these reasons, Australia is an attractive place to train AI. A more clearly circumscribed copyright framework can create operational advantages for organisations that design systems with compliance in mind.

Transparency, traceability and audit readiness

Where organisations are required to identify and document the source of training data, this can lead to stronger data provenance practices. Systems that incorporate dataset logging, version control and rights tracking are more readily auditable. Being able to demonstrate where data originates, what rights attach to it, and how it has been used can support regulatory engagement, due diligence and trust with customers and partners.

Commercial certainty and alignment with emerging standards

A licence based approach allows organisations to allocate risk more effectively and structure agreements around known rights, risks and permitted uses, rather than relying on legal arguments under broader doctrines of fair and transformative use (for example). This promotes more predictable cost modelling and reduces the incidence of post deployment disputes. It also aligns with a broader global shift towards transparency and accountability in AI systems, positioning organisations to meet evolving expectations across jurisdictions.

Key takeaways

Australia’s approach to copyright and AI training places a clear emphasis on where and how training activities take place. As AI infrastructure expands locally, these considerations arise at the point of system design and operation, not only after deployment.

For organisations building or deploying AI systems:

• infrastructure location is important - where training occurs can determine the applicable copyright regime and whether copying acts constitute an infringement;
• copyright infringement risks may arise during training – consider whether copyright material is reproduced or stored without authorisation as part of an AI training process;
• auditability, data provenance and rights tracking are core elements of compliant AI development;
• reliance on foreign laws and doctrines does not resolve these issues under Australian law.

Taken together, these factors reflect a clear and structured treatment of copyright works and approaches to AI training in Australia. For organisations willing to engage with the Australian approach, risk management may be more predictable and quantifiable, governance is clearer and AI systems more defensible. This positions Australia as a practical and credible environment for AI development.