As artificial intelligence becomes increasingly central to innovation, companies operating in Australia must navigate a complex legal landscape, especially when training AI systems on material in which others own copyright. While Australia has yet to see a court case directly addressing the use of copyright material to train an AI system, recent developments in the US and EU offer insight as to what the future might hold in this area.
In Australia, copyright protection extends to works protected in countries that are parties to the Berne Convention for the Protection of Literary and Artistic Works, first adopted in 1886 (Berne Convention). If companies infringe copyright in other Berne Convention jurisdictions such as the US and EU by training an AI system on copyright works where relevant defences do not apply, subsequent claims may be brought in Australia under sections 132AD – 132AM of the Copyright Act 1968 (Cth) (Copyright Act) for the importation or exploitation of an infringing copy of the AI system in Australia in certain circumstances. Under section 10 of the Copyright Act, an infringing copy includes an electronic copy of a work, the making of which would have constituted an infringement of copyright if the article had been made in Australia by the importer. However, sections 10AB and 44E of the Copyright Act provide that a copy of a computer program will be non-infringing where it was made in a Berne Convention country and did not infringe copyright in that country.
Fair dealing vs fair use: A narrower path in Australia compared to the US
In Australia, fair dealing defences under Part 3 of the Copyright Act allow reproduction only for specific purposes, such as research, criticism, parody, or accessibility, in certain circumstances. Under sections 43A and 43B of the Copyright Act, there are also fair dealing defences for temporary reproductions made in the course of communication, or as part of a technical process of use although Australian courts have not yet dealt with these provisions in the context of AI systems. The narrow scope of fair dealing defences means that companies must be especially cautious when training or deploying AI systems in Australia
Unlike the Australian law, US copyright law includes a broad fair use doctrine, meaning a greater variety of reproductions can be classified as non-infringing. Despite a broader definition of fair use, both jurisdictions assess fairness in relation to similar factors, including the purpose and character of the use, the nature of the work, the amount used and the market impact of the use.
Lessons from the US Copyright Office
In May 2025, the US Copyright Office released a pre-publication version of Part 3 of its Report on Copyright and Artificial Intelligence (US Copyright Office Report), which explores how generative AI systems use copyrighted works during training and deployment.
Copyright ownership carries with it exclusive rights including the exclusive right to reproduce a substantial part of a copyright work (subject to fair dealing and other defences).
The US Copyright Office Report considered the key question of whether training or use of generative AI models involves reproduction of copyright works so as to potentially give rise to a cause of action for copyright infringement. It identifies several ways reproduction of copyright works may occur, including:
- through the downloading, copying, and modification of copyright works for training datasets;1 and
- through the process of an AI system memorising copyrighted works and making temporary copies during model optimisation;2 and
- when an AI system produces “material that replicates or closely resembles copyrighted works”.3
The US Copyright Office Report Findings
The US Copyright Office Report provides a useful summary of key US principles. An overview of some of the key themes is set out below, followed by some recent cases.
Purpose and character of use
The US Copyright Office Report noted that under US law, AI training is more likely to be considered fair use when it is transformative, meaning it changes the original work significantly. For example, the report noted that “training a generative AI foundation model on a large and diverse dataset will often be transformative” since the large datasets are converted into statistical models which can produce a variety of outputs.4 However, if the AI is designed to replicate or compete with the original work (eg using Retrieval-Augmented Generation to serve the same purpose), it is less likely to be transformative and therefore less likely to be fair use.5 Importantly, the US Copyright Office Report notes that the AI training process is not “inherently transformative”, and the transformativeness of the use will depend upon the nature of the training process.66
Nature of the work
The US Copyright Office noted that “the use of more creative or expressive works (such as novels, movies, art, or music) is less likely to be fair use than use of factual works (such as computer code)” in training an AI system.7 The amount of a work used will also be relevant.8
Effect on the market
AI-generated content can dilute markets for original works by increasing competition and making it harder for audiences to find authentic content. If an AI output competes directly with the original work and results in “market substitution”, this weighs heavily against fair use.9 Additionally, according to the report, “where licensing markets are available to meet AI training needs, unlicensed uses will be disfavoured.”10
US case law: Implications for businesses
Recent US cases also offer practical insights:
• Thomson Reuters v ROSS Intelligence:11 A fair use defence was found not to be available when ROSS used Westlaw headnotes to train a legal AI tool. In this case, a team of humans wrote memoranda based on Westlaw headnotes to train the AI tool, and the majority of the memoranda were found to infringe copyright in the headnotes on which they were based. The court found the use non-transformative and commercially competitive. The case highlights the risk of using proprietary content for similar purposes under US law.
• Bartz v Anthropic:12 The court found that training Claude AI on books which it had legally purchased was fair use due to its transformative nature and lack of infringing outputs. However, the defence was found not to be available in relation to pirated books.
• Kadrey v Meta:13 Authors who alleged that copyright in their books had been infringed were found not to have established reproduction of their works nor any adverse effect on the market for them.
A number of lawsuits are continuing around the world, and precedents will continue to emerge over coming years. For example, Disney, Universal and others have sued Midjourney in the District Court of the Central District of California alleging infringement of copyright in popular characters and other works in its image generating service. That case is ongoing, and its outcome will not be known for some time.
The EU’s approach: Copyright, transparency and liability in AI
The EU has adopted a comprehensive and binding framework for regulating the intersection of AI and copyright. This dual regime includes:
• The AI Act14, which imposes obligations on developers of general-purpose and high-risk AI systems to document training data, label AI-generated content, and implement safeguards against the use of copyrighted material in certain circumstances. These obligations are enforceable by EU national regulators and the newly created AI Office, with penalties of up to 7% of global turnover for serious non-compliance.15
• The Copyright in the Digital Single Market Directive16, which provides two exceptions for text and data mining (TDM):
o Article 3 permits non-profit research organisations and cultural institutions to conduct TDM for scientific research, provided lawful access. This exception is mandatory and cannot be overridden by contract.17
o Article 4 allows commercial and general-purpose TDM uses, including for AI training, unless the rights holder has opted out in an appropriate manner, such as by way of machine-readable means (eg metadata, robots.txt, or clear website terms) for material which is publicly available online.18
Developers training AI on EU-based content must verify whether a valid opt-out applies and secure licences where required. Failure to do so may result in copyright infringement.19
This framework was applied in a 2024 landmark decision by the Hamburg Regional Court where LAION, a non-profit organisation, had compiled an AI training dataset consisting of hyperlinks and metadata associated with images available online.20 The Court upheld the application of the scientific TDM exception under German law (transposing Article 3 of Directive (EU) 2019/790), finding the dataset creation to fall within research purposes – even where future commercial use by third parties was possible. The Court also accepted that opt-outs may be validly expressed in natural language, provided they are machine-readable at the relevant time.
Under EU law, AI-generated content is not eligible for copyright protection unless it reflects a meaningful human creative contribution. This requirement of human authorship is consistent with the principles of the Berne Convention.21
In addition, the Product Liability Directive applies in certain circumstances in relation to AI systems that are defective or unsafe.22
AI regulation in Australia
Last year, the Final Report of the Australian Senate Select Committee on Adopting Artificial Intelligence recommended a series of reforms and measures to address the use of copyright works by AI systems.23 Specifically, it recommended that that the Government require developers of AI products to be transparent in relation to their use of copyrighted words in training datasets and ensure that the use of such works is appropriately licensed and paid for.24 The report also advocated for whole-of-economy legislation to regulate high-risk AI applications,25 similar to the EU’s AI Act.
More recently, the Productivity Commission’s report on ‘Harnessing data and digital technology’ recommended that copyright issues posed by AI should be resolved through existing copyright law frameworks rather than through the introduction of specialist AI legislation.26 This was contrary to the recommendation of the Senate Select Committee’s recommendation for whole-of-economy legislation to regulate high-risk AI applications.27
In relation to the training of AI systems, the Productivity Commission has noted that jurisdictions such as the EU have text and data mining fair dealing exceptions in their copyright laws and seeks views on whether such exceptions should be introduced in Australia.28 Such an exception would cover “not just AI model training but all forms of analytical techniques that use machine-read material to identify patterns, trends and other useful information.”29 The Productivity Commission seeks views as to whether a text and data mining exception should be introduced in the Copyright Act, and if so, whether it should be limited to non-commercial uses, and whether any additional criteria or regulatory guidance should be included.30 Notably, the Minister for Industry and Innovation and Minister for Science, the Hon. Tim Ayers has indicated that the Government has no plans to introduce a text and data mining exception.31
The coming year is expected to be pivotal in shaping Australia’s regulatory landscape for AI. Continued consultation with creative industry stakeholders, rightsholders, and their representative organisations will be important to develop fair and transparent mechanisms for licensing and compensation.32
Strategic takeaways for companies operating in Australia
It is important for any companies developing, importing or training AI software in Australia to be cognisant of copyright issues. Care should be taken to understand and appropriately manage relevant risks. Copyright owners should also be aware of the possibility that their copyright works (and other material in which copyright subsists, such as sound recordings and cinematograph films) may be used to train third party AI. Measures such as paywalls, measures to block AI bots, or prohibitions in applicable terms and conditions can be used to manage this risk where the commercial balance weighs in favour of restricting such copying.
- US Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training (Prepublication Copy, 9 May 2025) (US Copyright Office Report) pp 26-28. ↩︎
- US Copyright Office Report pp 28-30. ↩︎
- US Copyright Office Report p 31. ↩︎
- US Copyright Office Report p 45. ↩︎
- US Copyright Office Report pp 46-47. ↩︎
- US Copyright Office Report pp 47-48. ↩︎
- US Copyright Office Report p 53. ↩︎
- US Copyright Office Report p 55. ↩︎
- US Copyright Office Report pp 62, 73. ↩︎
- US Copyright Office Report p 71. ↩︎
- Thomson Reuters Enterprise Centre GmbH v ROSS Intelligence Inc (D Del, No. 1:20-cv-613-SB, 11 February 2025). ↩︎
- Bartz v Anthropic PBC (ND Cal, No 24-cv-05417 23 June 2025). ↩︎
- Kadrey v Meta Platforms Inc (ND Cal, No 23-cv-03417, 25 June 2025) ↩︎
- Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024. ↩︎
- European AI Act (Regulation (EU) 2024/1689), art 99, para 3. ↩︎
- Directive (EU) 2019/790 of the European Parliament and of the Council of 17 April 2019 on copyright and related rights in the Digital Single Market and amending Directives 96/9/EC and 2001/29/EC (Directive (EU) 2019/790) ↩︎
- Directive (EU) 2019/790 art. 4. ↩︎
- Directive (EU) 2019/790 art. 3. ↩︎
- Directive (EU) 2019/790 art. 4, para 3. ↩︎
- Hamburg District Court, 310 O.227/23, LAION v Robert Kneschke. ↩︎
- Berne Convention for the Protection of Literary and Artistic Works (as amended on September 28, 1979). ↩︎
- Directive (EU) 2024/2853 of the European Parliament and of the Council of 23 October 2024 on liability for defective products and repealing Council Directive 85/374/EEC. ↩︎
- The Senate, Select Committee on Adopting Artificial Intelligence (Final Report, 2024) (Senate Committee Report) ↩︎
- Senate Committee Report p 16. ↩︎
- Senate Committee Report p 15. ↩︎
- Australian Government Productivity Commission, Harnessing data and digital technology (Interim report, 2025) (Interim Report) p 9. ↩︎
- Senate Committee Report p 15. ↩︎
- Interim Report p 24 ↩︎
- Interim Report p 27. ↩︎
- Interim Report p 27. ↩︎
- Copyright Agency, Copyright Agency welcomes Government’s continued support for Australia’s copyright system in the age of AI (Web Page, 22 August 2025)
<https://www.copyright.com.au/2025/08/copyright-agency-welcomes-governments-continued-support-for-australias-copyright-system-in-the-age-of-ai/>. ↩︎ - Senate Committee Report p 16 ↩︎
