Most of the public debate about AI and copyright has focused on whether scraping the internet to train a model counts as infringement, and whether developers should pay licensing fees to content creators. These are important questions. But there is a more fundamental question sitting underneath all of them that almost nobody is asking: who actually owns the trained AI model itself?
For Australian founders and business owners who rely on AI tools, this is not just an academic puzzle. If the ownership of these models is legally contested, the licences you are paying for may be built on shakier ground than you think. Here is what you need to understand.
The Omelette Problem
There is a useful way to think about how frontier AI models were built. Imagine that every person who has ever written something, taken a photograph, drawn an illustration, written code, or created any other content that ended up on the internet owns one egg. A developer arrives, takes every egg without asking or paying, adds a few eggs of their own (the training code, the fine-tuning data, the evaluation tools), scrambles everything together into the largest omelette ever made, and then announces they are the sole owner. They then charge the original egg-owners to eat it.
This is, in plain terms, how frontier AI models were constructed. The technical language around weights, parameters, vectors, and embeddings can make it feel abstract, but the underlying fact pattern is one the law has dealt with for centuries: what happens when you mix together property belonging to many different people and end up with something new?
Copyright Attaches Automatically in Australia
Under Australian law, copyright attaches to a literary, artistic, musical, or dramatic work the moment it is created and fixed in some form. The author owns it automatically. No registration required. This principle flows from the Copyright Act 1968 (Cth) and Australia's obligations under the Berne Convention.
Critically, someone else using your work, even transforming it significantly, does not extinguish your ownership of what you originally made. Running a literary work through a GPU does not destroy the original author's copyright any more than scanning a document destroys the original.
AI developers building and monetising frontier models are implicitly relying on two premises that are not found anywhere in Australian copyright law:
- First: that ingesting a work into a training process extinguishes the original author's ownership interest in what was contributed.
- Second: that this extinguished interest then re-allocates to the developer rather than to the public domain or to no one.
Neither premise is stated explicitly by developers, and neither is supported by the Copyright Act 1968, the Berne Convention, or any decided case. They are simply assumed.
Why Standard Copyright Remedies Do Not Work Here
Copyright law was designed with something like a photocopier in mind. An original goes in, copies come out, and the remedies available, injunctions, damages, delivery up of copies, accounts of profits, operate on discrete, countable reproductions.
A trained AI model does not work this way. The works fed into training do not come out the other side as recognisable copies. They are metabolised into weights and parameters that shape every output but are not themselves recoverable. There is no copy to impound or destroy. This creates a genuine gap in the law:
- An injunction means little once the model already exists. Cease and desist what, exactly?
- Delivery up of copies is not available because the model is not a copy and cannot be decomposed back into its constituent works.
- An account of profits is largely moot because the major frontier AI developers currently operate at significant losses and are funded by venture capital rather than commercial revenue.
The standard copyright toolkit simply does not reach the asset that was actually built from everyone's work. This is why the ownership question matters so much: it has to be resolved upstream, before infringement analysis even begins.
What the Australian Government's Response Tells Us
In December 2023, the Australian Government established the Copyright and Artificial Intelligence Reference Group (CAIRG) to examine these issues. The group has framed the problem primarily as a licensing question, meaning it is working toward a system where AI developers pay for access to training data.
But there is an important concession buried in that framing. A licence only makes sense if the licensor owns something the licensee does not already have a right to. If a licensing scheme is necessary, it implicitly acknowledges that developers do not already own the entirety of what went into making their models. That is an admission, not a solution. It brings you straight back to the collective ownership question without answering it.
What This Means for Your Business
If your business relies on a frontier AI tool, whether for customer service, content creation, legal drafting, code generation, or any other function, there are two practical implications worth keeping in mind.
Your AI licences are based on possession, not clear title. When you pay an AI API provider, you are paying a party whose claim to own what they are licensing rests largely on the fact that they built and control access to the model. The legal underpinning for that ownership claim is, at best, untested and, at worst, wrong. That does not mean you should stop using these tools. It means you should understand what you are actually paying for and avoid building critical business infrastructure on the assumption that the current ownership arrangements are settled.
Data governance matters more than ever. Regardless of how the ownership debate resolves, businesses handling customer data or feeding proprietary information into AI systems need to think carefully about their data practices. A well-drafted Privacy Policy is not just a compliance formality. It is how you document what data you collect, how it is used, and what your obligations are to the people whose information you hold. If your business is using AI tools that process customer data, your privacy documentation needs to reflect that.
Confidentiality is a live issue. If you are sharing proprietary business information, client data, or commercially sensitive material with AI platforms, consider whether your existing agreements adequately address this. A Confidentiality Agreement with contractors, partners, or employees who use AI tools in their work is one practical way to manage the risk of inadvertent disclosure.
The Bigger Picture
The conservative legal position, the one that existing law arguably already supports, points toward collective ownership of frontier AI models. Under the confusion doctrine, when property belonging to many parties is mixed together inseparably, the resulting asset is owned collectively in proportion to contribution. That would mean, in principle, that every person whose work was used in training has some ownership stake in the resulting model, alongside the developer.
There is precedent for this kind of institutional arrangement. CERN, the European Southern Observatory, and NATO infrastructure have all used common-funded contractor models where a shared asset is maintained collectively and contractors build and operate it without holding sole title. The subject matter here is new. The structure is not.
The alternative is sole developer ownership, but that requires accepting two premises the law does not supply and that would effectively strip every living author of rights the Berne Convention was designed to protect.
There is no comfortable middle ground. As this debate moves toward resolution, whether through legislation, international agreement, or litigation, the outcome will have real consequences for every business that has built workflows around AI tools.
If you are reviewing your business's legal foundations in light of AI adoption, Mode.law's document library at /documents includes practical templates for privacy policies, confidentiality agreements, and other documents relevant to technology businesses. Good documentation will not resolve the ownership debate, but it will put your business in a stronger position whatever the outcome turns out to be.