Copyright in AI-generated content in Australia and who owns ChatGPT output

In Australia, "copyright free" is a misleading label for ChatGPT output. Copyright protects original works created through human intellectual effort, so purely machine-generated text may not attract copyright protection at all. But an absence of protection is not a licence to use: publishing AI output can still infringe someone else's copyright, and the provider's contract terms, privacy law and consumer law obligations all continue to apply.

This article explains what copyright in AI-generated content actually means for your business:

• Subsistence: what copyright protects, and why it starts with an original human author
• The human authorship rule: why purely AI-generated output may be unprotected, and how human input changes the result
• Contract terms: what the provider's terms actually give you, and what they cannot give you
• Infringement: when publishing AI output can still breach someone else's copyright
• Practice: a worked example, the common misconceptions, and when legal help is worth getting

What copyright protects: original works, created automatically

Under the Copyright Act 1968 (Cth) (the Act), copyright subsists in original literary, dramatic, musical and artistic works (s 32). Protection is automatic. Australia has no registration system for copyright, unlike trade marks and patents. The moment a work is written down, recorded or otherwise made in material form, copyright comes into existence without any paperwork or fee.

Copyright gives its owner a bundle of exclusive rights, including the right to reproduce the work, publish it and communicate it to the public. The Act's infringement rule is correspondingly simple: a person who, not being the owner, does any act comprised in the copyright without the owner's licence infringes it (s 36).

The decisive word in s 32 is "original". Australian courts have held that originality does not mean novelty or merit. It means the work originates from its author's independent intellectual effort. In IceTV Pty Ltd v Nine Network Australia Pty Ltd ([2009] HCA 14), the High Court confirmed that a work is original where the author has exercised skill, judgment and effort in producing the expression of the work. That threshold is deliberately low, but it is real, and it is where AI output starts to get difficult.

The human authorship rule

Copyright in Australia is built around human authors. The Act does not define "author" to include a machine, and the case law makes clear that a work must originate with a human being.

The leading authority is Telstra Corporation Ltd v Phone Directories Company Pty Ltd ([2010] FCAFC 149). The Full Federal Court held that copyright did not subsist in telephone directories that were compiled with substantial automation, because no identifiable human had exercised the necessary intellectual effort in creating the final form of the work. The reasoning translates directly to generative AI: when a model produces text, no human contributed independent intellectual effort to the expression itself, so the output may not qualify for protection as a literary work.

Australian IP law treats machines the same way elsewhere. In Commissioner of Patents v Thaler ([2022] FCAFC 62), the Full Federal Court confirmed that an AI system could not be an inventor under the Patents Act 1990 (Cth), because an inventor must be a natural person. Australian copyright law also has no equivalent of the UK rule that treats the person who makes the arrangements for a computer-generated work as its author. The Australian Government is still working through these questions through its Copyright and Artificial Intelligence Reference Group, established in December 2023, but as at the date of writing the Act has not been amended to recognise AI as an author.

The analysis is not binary, however. Human authorship exists on a spectrum. If a person contributes original expression through detailed prompting, restructuring, editing and curating the output, the final published work may well contain human-authored elements and attract copyright in those elements. The more meaningful and original the human contribution to the final text, the stronger the claim to protection. A business that keeps records of its prompts and edits is in a much better position to demonstrate that contribution if ownership is ever disputed.

One further consequence of the human authorship rule is worth knowing. Moral rights, which include the right of attribution and the right of integrity, attach only to human authors under Part IX of the Act. A machine has no moral rights, and the human who produces the final edited work does.

What the provider's terms actually give you

The provider's contract is often the first thing people point to when asking who owns ChatGPT output. OpenAI's Terms of Use provide that, as between you and OpenAI and to the extent permitted by applicable law, you retain your ownership rights in your input and you own the output, with OpenAI assigning to you all its right, title and interest, if any, in the output (OpenAI Terms of Use).

Three points follow from that wording:

• Permission to use: the terms give you contractual permission to use the output, including commercially, subject to the restrictions in the terms. That is genuinely useful: it resolves the question of whether the provider can stop you using what you generated.
• The limits of the assignment: the assignment operates "to the extent permitted by applicable law". A contract cannot manufacture copyright protection where the Act requires a human author. If the output is purely machine-generated with no meaningful human input, there may be no copyright for you to own in a legal sense, even though the terms say you "own" the output. What you have is a contractual right to use it, not a statutory monopoly over it.
• No protection against third parties: the terms do not protect you against third parties. If your output reproduces a substantial part of someone else's protected work, the fact that OpenAI assigned the output to you does not answer a claim by that third party. OpenAI also warns that output may not be unique and that other users may receive similar output, which matters if your plan is to build a distinctive brand voice on AI-generated text.

AI output can still infringe someone else's copyright

The infringement rules turn on what you do with protected material, not on whether a human or a machine produced the infringing text. Two provisions of the Act do most of the work:

• The infringement rule in s 36: it applies to anyone who does an act comprised in the copyright, such as reproduction, without the owner's licence. There is no exemption for content that was generated with an AI tool.
• Substantial part in s 14: s 14 of the Act provides that doing an act in relation to a substantial part of a work is treated as doing it in relation to the whole work. That matters because generative models are trained on large volumes of existing material and can reproduce passages of it, even though they are designed to produce new text.

"Substantial part" is a qualitative test, not a quantitative one. A short passage can be substantial if it takes the heart of the work, such as its distinctive phrasing, structure or selection and arrangement. The practical risk areas for a business publishing AI-assisted content include:

• Verbatim passages: output that closely mirrors an identifiable published source is high risk, whether or not you intended it.
• Distinctive expression: copying a unique turn of phrase or the structure of a piece can infringe even where the amount copied is small.
• Prompted copying: pasting chunks of third-party articles, client materials or paid reports into a prompt and then publishing all or part of that text can infringe copyright or breach confidentiality obligations.

What about asking the AI to write "in the style of" a particular author or brand? Copyright protects expression, not ideas or general styles, so style imitation alone is less risky. But if the output ends up closely matching a recognisable protected text, the style instruction is no defence.

Fair dealing exceptions will rarely rescue routine business publishing. The Act allows fair dealing only for narrow, enumerated purposes: research or study, criticism or review, parody or satire, reporting news, and professional advice given by a lawyer or patent attorney. Commercial marketing, product descriptions and general business publishing do not fit these categories, so it is unwise to assume an exception applies to AI-generated copy.

How this plays out: a worked example

Consider a boutique skincare business in Brisbane preparing to launch a new product range. The founder asks ChatGPT to draft product descriptions and a launch blog post, and pastes a competitor's press release into the prompt asking for a "rewrite".

Start with ownership. The first raw output is largely machine-generated. Under current Australian principles, that output likely lacks human authorship and therefore does not attract copyright. If a rival copied it, the founder would have little ability to sue for infringement.

Now change the facts slightly. The founder heavily edits the descriptions, adding original detail about formulations, ingredient provenance and the brand story. Those final texts are much more likely to be protected, because they now contain her independent intellectual effort. Her claim extends to the human-authored elements of the final work, even if the machine-generated starting point was unprotected.

The infringement risk remains separate. The launch post still contains a paragraph that tracks the competitor's press release almost verbatim. Because that paragraph is a substantial part of a protected work in a qualitative sense, publishing it could infringe the competitor's copyright. The instruction to "rewrite" the material does not change that analysis, and neither does the fact that an AI produced the draft.

There is a third layer in this example worth noticing. If the founder pastes her customer email list into the same tool to draft a marketing email, she raises issues under the Privacy Act 1988 (Cth) and her confidentiality obligations that have nothing to do with copyright. Copyright is the concept in question here, but it is rarely the only legal issue in an AI workflow.

Common misconceptions

Several misconceptions about AI output and copyright are worth naming specifically, because each one leads to a different kind of mistake:

• "AI content is copyright free, so anyone can copy it": Wrong. A lack of protection for the AI-generated material does not make it free to use. You cannot stop others from copying your raw AI output, and you can be liable if your output reproduces a substantial part of someone else's work.
• "OpenAI owns everything I generate": Wrong. OpenAI's terms assign the output to you. The real limitation is legal, not contractual: without human authorship there may be no copyright to own.
• "I generated it, so I own it and can stop others using it": Wrong where the output is purely machine-generated. Copyright protection in Australia depends on human authorship, so there may be nothing to enforce.
• "Changing the words a bit makes it safe": Wrong. The substantial part test is qualitative, and a lightly edited reproduction of distinctive expression can still infringe.
• "Fair dealing covers business use": Wrong. The exceptions are narrow and enumerated, and commercial publishing rarely fits within them.

Getting help: what a copyright lawyer actually does

A lawyer's job in this area is to work out where your particular workflow sits on the human authorship spectrum, and to put guardrails around it so that legal risk does not slow the team down. Concretely, that usually means reviewing how you prompt, edit and record your AI use so you can demonstrate human authorship if ownership is challenged; checking whether the provider's terms and any third-party licences cover what you plan to do; running clearance checks on high-value assets such as taglines, hero pages and campaigns; and advising on whether fair dealing or any other exception applies to a specific use.

The same practitioner will typically draft the documents that make the position enforceable across the business: a generative AI use policy, supplier and contractor agreements with warranties that deliverables are original and do not infringe third-party rights, non-disclosure agreements where confidential material might find its way into prompts, and website terms and conditions covering how AI-assisted content is published. Where AI output touches personal information or makes claims about products, the advice will also extend to privacy obligations and misleading or deceptive conduct under the Australian Consumer Law.

The benefit of this review is practical. A team that knows which content is protectable, which uses are safe and which documents govern its AI tools can publish with confidence, instead of discovering the gaps after a takedown notice or an infringement claim arrives.

The question to ask before you publish

Before any AI-assisted text goes live, ask two questions. Would a court see a human author behind the final expression, or is this essentially machine output? And does any of it reproduce an identifiable protected source, even briefly?

If you cannot answer both comfortably, the content is not "copyright free". It is unprotected and potentially infringing at the same time, which is the worst of both worlds. That combination is exactly when the cost of a quick legal review is trivial compared with the cost of defending an infringement claim, or of discovering that a rival can freely copy the content you built your brand around.