This is part 5 of our series about ‘Law in the AI age’ explores the Attorney-General’s recent rejection of including an AI training exception in the Copyright Act. Below we analyse what it actually means: Artist watershed, toothless tiger or self-sabotage?
On 27 October 2025, Australian Attorney-General Michelle Rowland, speaking on ABC radio, explicitly rejected introducing a text and data mining exception into the copyright law (TDM Exception).
This comes after the Government Productivity Commission’s Report explored introducing a TDM exception in the context of an inquiry into the opportunities and challenges presented by AI in Australia. In response, our recent article had taken the position that a TDM Exception should be introduced, though with guardrails.
Below we explore some practical consequences of rejecting a TDM Exception. What does this actually mean for Australian copyright-owners and AI platform owners? Spoiler: It doesn’t stop ChatGPT from legally training on Australian artists’ copyright-protected works.
Following on from the world’s first ban for teens on social media, Australia’s legal position in the tech landscape has recently displayed an individualistic flair.
Other jurisdictions do have some form of a TDM exception. For example, the EU’s AI Act allows AI models to train on copyrighted materials for commercial purposes, unless the copyright-owner has opted out. Meanwhile, Japan has a unique model that allows AI models to train on copyrighted materials if it is for “non-enjoyment” purposes; that is, the aim of the training cannot be to enable a person to merely “enjoy yourself!” with the work.
Without a TDM exception, it is likely that AI models that train on copyrighted content in Australia infringe copyright. As we covered in our recent ‘fair play’ article, this is because AI training means that the AI model to at least temporarily reproduce the works, which is a right exclusively reserved to the copyright-owner.
The Copyright Act is a territorial Act confined to Australian shores. To breach the Act, the infringing conduct must occur within Australia.[1] This is unlike the Privacy Act, which has express extra-territorial operation where the infringer meets the test of carrying on business in Australia.
The US is one of the world’s biggest exporters of AI, and OpenAI’s ChatGPT is the world’s most-used AI tool by an overwhelming margin, having recently reached 800 million weekly active users.[2] The major AI models are all global companies operating internationally, including the following models backed by major US media companies:
Whilst China’s DeepSeek could also be added to the list, the point is that training of these AI models is occurring off Australian shores.[3]
Where the rubber hits the road is that an aggrieved Australian copyright-owner typically needs to enforce their rights against the US models under US copyright law, not under the Australian Copyright Act. This arises from the ‘national treatment’ principle of the Berne Convention’,[4] of which most countries, including Australia, the US and China, are signatories. In practical terms, this means that if an Australian person creates a copyright-protected literary work in Australia, and it is later published in the US, then US copyright law will govern the author’s rights.
In the US, the ‘fair use’ exception is a key defence for tech companies that train AI models on copyright-protected works. A key factor is whether the use is ‘transformative’; that is, whether it adds ‘something new, with a further purpose or different character, altering the first with new expression, meaning, or message.’[5]
The US fair use doctrine is much broader than Australia’s ‘fair dealing’ exceptions, which only apply in narrow and prescribed circumstances.
This results in a judicially-driven model to copyright protection that relies on case-by-case interpretation. For example, in June 2025, Meta’s use of authors’ books to train its AI model LLaMa was found to be ‘fair use’ on the basis that the authors had not presented enough evidence that the AI model would cause ‘market dilution’ by flooding the market with work similar to theirs.[6] More recently, Disney and Universal sued Midjourney alleging copyright infringement in popular characters in its image generating service. There are various lawsuits on foot, and this is a continually developing legal landscape that will be interesting to monitor.
The fair use doctrine contrasts with the TDM Exceptions adopted in the EU and Japan, which take a statutorily prescriptive approach to copyright protection.
Without a TDM exception or US-type fair use protection, Australian AI now has a head-on issue; how to legally train and operate their AI models in Australia. Already faced with the difficult challenge of taking on US competitors from down under, it’s a belly blow for Aussie AI. It undoubtedly creates a stronger incentive for AI tech founders to migrate their businesses offshore, as if there was not already a well-worn ‘yellow brick road’ to California; a road paved with gold bricks by the gargantuan Venture Capital industry.
At least Australian users of AI models can breathe safe; they are unlikely to be in breach of Australian copyright law. This is because, as a mere passive consumer of the AI output, the user does not infringe copyright law.
To balance the interests of the Australian tech industry and Australian copyright-owners, we recommend the urgent introduction of a collective licensing scheme that provides tech companies with clear authority to legally train on copyrighted materials, while ensuring fair remuneration for Australian copyright holders. As discussed in our recent Response to the Government’s Copyright Inquiry, the APRA/AMCOS model for music rights offers a useful model, where creators receive distributions based on AI usage share.
Others believe that the market will naturally regulate itself. As stated by ARIA and PPCA CEO, Annabelle Herd, ‘hundreds of AI licensing deals are already operating at commercial scale, from Merlin’s licence with ElevenLabs to OpenAI’s agreements with News Corp and The Guardian. The fact that these deals already exist exposes the myth that the current law is somehow uncertain or not fit for purpose.’
However, an awkward question hangs in the air as to what incentive the dominant AI players have to make deals with Australian copyright-owners. They are not subject to the Australian Copyright Act, and can access the US fair use defence.
Is taking this stance on balance good for Australia? One can understand and in many ways applaud the driver to champion Australian creative artist rights. However, on a macro view, the decision to not implement a TDM exception primarily affects only Australian-based AI models, and this when every dominant AI player is already not Australian based.
Former Atlassian CEO, Scott Farquhar, has said that a TDM Exception ‘could unlock billions of dollars of foreign investment into Australia’. The absence of a TDM Exception risks isolating the Australian tech industry from the global AI economy, while in truth does little to protect the works of Australian copyright-owners.
Speaking practically, how will approvals be given and compensation exchanged? Or will Australian artists be ignored globally, as offshore AI models continue to legally train on their copyright-protected works without authorisation?
Australia has recently moved to a unique regulatory position in a global landscape, highlighted by the under 16s social media ban, and now the Attorney-General’s rejection of a TDM Exception. The ‘rebel without a clause’, Australian Attorney-General Michelle Rowland has said, “There is a body of work to do around what the copyright environment looks like in the AI world”. Much more on this topic is yet to be mined, but the pressure is building downunder.
Does your business use artificial intelligence? For specific advice on how to use and commercialise AI in your business, please contact us below. Edwards + Co Legal provide corporate and commercial legal advice to modern Australian businesses.
[1] Copyright Act, s 36(1).
[2] https://techcrunch.com/2025/10/06/sam-altman-says-chatgpt-has-hit-800m-weekly-active-users/.
[3] OpenAI’s trains on three primary sources of information: (1) information that is publicly available on the internet, (2) information OpenAI partners with third parties to access, and (3) information that OpenAI’s users, human trainers, and researchers provide or generate. https://help.openai.com/en/articles/7842364-how-chatgpt-and-our-foundation-models-are-developed
[4] The ‘national treatment’ principle provides that works originating in one of the member states must be given the same protection in each of the other member states as the latter grants to the works of its own nationals.
[5] https://www.alrc.gov.au/publication/copyright-and-the-digital-economy-ip-42/transformative-use/.
[6] https://www.theguardian.com/technology/2025/jun/26/meta-wins-ai-copyright-lawsuit-as-us-judge-rules-against-authors.
