- Who is involved in the AI copyright picture
- Stage one: protection, originality and the human author
- Stage two: who owns what
- Stage three: how infringement happens
- Stage four: the rights that do not depend on copyright
- Where the law is unsettled
- Where a lawyer earns their fee
- The evidence question to answer before you publish
Most Australian businesses now have an AI tool somewhere in the content pipeline. Copy for a product page, a logo draft, a set of social posts, a first cut of code: the output arrives in seconds and looks finished. The legal question is rarely asked until something goes wrong. What does copyright do when the creative work was done by a machine?
Australian copyright law is a statutory scheme built on a simple assumption. Every protected work has a human author. The Copyright Act 1968 (Cth) (the Act) protects original literary, dramatic, musical and artistic works, and the entire structure of ownership, duration and moral rights hangs off that person. Artificial intelligence does not fit the assumption comfortably, which is why the answers below are expressed as probabilities rather than certainties.
This article walks through the scheme as it operates today: what triggers protection, who ends up owning what, how infringement happens even when an AI tool did the copying, and the rights that protect you regardless of copyright. It becomes relevant at the moment you are about to publish AI-assisted material, launch an AI-generated brand identity, or feed proprietary content into a tool.
Who is involved in the AI copyright picture
Five groups have an interest in how copyright applies to AI output:
- Your business: creates and publishes the material, and either owns the rights or needs permission to use it.
- The AI provider: supplies the tool, sets the terms governing your inputs and outputs, and trains models on large datasets.
- Human creators and rights holders: the authors whose works sit in training data or licensed libraries, and who hold moral rights that survive licensing.
- Competitors: can pursue you under trade mark law, passing off or consumer law even where copyright in an output is unclear.
- Courts and Parliament: courts apply the Act as written; Parliament has not yet amended it for AI.
Each of these players has a different interest, and the scheme allocates rights and risks between them according to the stages set out below.
Stage one: protection, originality and the human author
The trigger for protection is simple to state and hard to apply. Under s 32 of the Act, copyright subsists in an original literary, dramatic, musical or artistic work made by a qualified person. Protection is automatic. There is no register, no application and no fee; in Australia you do not file for copyright the way you file a trade mark or patent. The work simply needs to be original, and it needs an author.
Originality does not mean novelty or artistic merit. The High Court held in IceTV Pty Ltd v Nine Network Australia Pty Ltd [2009] HCA 14 that a work is original when it originates from its author through independent intellectual effort, and that the effort must go into the form of expression rather than just the underlying information. Copyright protects the way ideas are expressed, not the ideas themselves.
The Full Court of the Federal Court applied that standard to automated output in Telstra Corporation Ltd v Phone Directories Co Pty Ltd [2010] FCAFC 149. Telstra's telephone directories were compiled largely by an automated computer process, and the Court held that copyright did not subsist in them: copyright in a literary work can subsist only if it originates from an individual, and machine-generated output without an identifiable human author fell outside the scheme.
That is the authority on which the AI analysis rests. No Australian court has yet decided squarely whether AI-generated content can be protected, so the position is reasoned by analogy: the less a human did, the less likely copyright exists. A single prompt such as "write a tagline for our café" will rarely make you the author of whatever comes back, because the model supplied the expression. But detailed direction, several rounds of prompting, selection between candidate outputs, and substantial editing or arrangement may tip parts of the final work into protectable territory, and typically only to the extent of the human contribution.
The Act has not been updated to add a special rule for computer-generated works, the way some overseas jurisdictions have done. As of the current compilation of the Act (April 2026) there is no provision on AI authorship, no provision on training models on copyrighted material, and no text-and-data-mining exception. The law is what the cases make it, and the cases point toward human authorship as the gate.
Stage two: who owns what
Ownership follows authorship as the default. Under s 35 of the Act, the author of a work is the first owner of the copyright, subject to exceptions. The important exception for a business is employment: if an employee makes a work under a contract of service in the course of employment, the employer owns the copyright. Section 35 also allows the default rules to be varied by agreement, so contracts can and should deal with ownership expressly. Commissioning a designer or developer does not, by itself, transfer copyright; for the human-authored parts you need an assignment or a clear IP clause in the contract.
Now apply the AI layer. If an employee writes copy with heavy AI assistance, the human contribution is still the employee's work, so the employer can hold copyright in those human-authored elements in the usual way. If a contractor produces an AI-assisted logo, ownership of the human contribution sits with the contractor unless the contract assigns it, and any purely machine-generated element may not be protected at all.
Two questions are often run together and should be kept apart. Is the output protected? That is a question about the Act. Who may use it? That is largely a question of contract. Platform terms that say you "own" outputs, or that grant you a licence to use them, create rights between you and the provider; they do not manufacture copyright where the Act does not confer it, and they bind only the parties to the contract. A licence from the provider does not stop a third-party rights holder from suing you if the output copies their work.
The practical consequence is to keep a record of what a human actually did. Save drafts, note your prompts and the choices you made between outputs, keep the edit history. With no registration system, that trail is your evidence of authorship if you ever need to enforce rights or defend a challenge.
Stage three: how infringement happens
Protection is the half of the scheme that helps you. Infringement is the half that can hurt you, and it operates whether or not the AI output is itself protected.
Under s 36 of the Act, a person infringes copyright by doing in Australia, or authorising the doing in Australia of, any act comprised in the copyright, without the licence of the owner. For literary works the exclusive rights include reproducing the work in a material form. And under s 14, taking a substantial part counts as taking the whole; substantiality is assessed by quality, not quantity, so even a short passage can be enough.
The consequence for AI use is direct. An output that closely echoes a protected song, image, code snippet or passage of text can infringe even if the similarity was produced inside the model and you never saw the source work. Intention is not an element of infringement. The fact that a tool did the copying does not give you a defence.
Training data raises the same questions one step earlier. Models are trained on large datasets that may include protected works, and the Act contains no exception for text and data mining of the kind used to train AI. Whether training itself infringes is not settled in Australia, and courts overseas have reached different conclusions. What is clear is the exposure on the output side, and it becomes your exposure the moment you publish the material.
Businesses that direct staff to use AI also sit in the territory of authorisation. Section 36(1A) lists the factors a court considers in deciding whether someone authorised an infringement: the extent of the person's power to prevent the act, the nature of the relationship with the person who did it, and whether reasonable steps were taken. A business that instructs a team to generate marketing content, or that builds an AI feature for its customers, should expect those factors to be examined.
Practical guardrails: treat AI output as a draft rather than a finished asset, run distinctive images through reverse image searches, run text and code through similarity or plagiarism checks, avoid prompts that ask for the style of a named artist or that reproduce a specific work, and have a named person review material before it goes live.
Stage four: the rights that do not depend on copyright
Copyright is not the only protection, and it is not the only risk.
Brand rights do not depend on authorship at all. A trade mark registered under the Trade Marks Act 1995 (Cth) protects a sign as a badge of origin, and an AI-generated logo can be registered if it is distinctive and not confusingly similar to an existing mark. Passing off protects the goodwill of a business against misrepresentation. The Australian Consumer Law, which sits in Schedule 2 to the Competition and Consumer Act 2010 (Cth), prohibits misleading or deceptive conduct in trade or commerce. Claims such as "powered by AI" must be accurate, and branding that is close to a competitor's can attract action under these heads even where copyright in the output is doubtful. Using AI does not shield you from any of this.
Moral rights are a separate layer that belongs to individuals only. Section 190 of the Act provides that only individuals have moral rights; a company cannot hold them, and they cannot be transferred with the copyright. Authors of works hold the right of attribution, the right not to have authorship falsely attributed, and the right of integrity, which is the right not to have the work subjected to derogatory treatment. An author can consent in writing to acts that would otherwise infringe, but the consent must be genuine and specific.
The practical point: if you hold a licence for stock images, music or code and then adapt them with AI, the licence deals with the copyright, but the moral rights stay with the human creator. Heavy reworking of licensed material can raise attribution and derogatory treatment issues, so check the licence, keep attributions, and be careful how far you transform licensed works.
Where the law is unsettled
The honest summary is that the copyright framework was not designed for AI, and Parliament has not yet updated it. As of the April 2026 compilation, the Act contains no AI-specific provisions.
The reform options under discussion in Australia and overseas give a sense of what may change: a statutory threshold for the human involvement needed to protect AI-assisted work, a copyright exception for using protected material to train models, and rules around digital replicas of a person's voice or likeness. None of it is law yet.
Two conclusions follow for a business. First, treat copyright in AI output as an uncertainty to be managed rather than an asset to be assumed. Second, the absence of copyright does not mean the material is free for the taking. Trade marks, passing off, consumer law, confidentiality and contract all continue to apply, and a business can still be liable for using or publishing AI output that steps on someone else's rights.
The same caution runs through the patent side. In Commissioner of Patents v Thaler [2022] FCAFC 62, the Full Court of the Federal Court held that an artificial intelligence system could not be named as an inventor under the Patents Act 1990 (Cth), because only a natural person can be an inventor. The reasoning is specific to the Patents Act, but it shows the direction Australian courts have taken: the statutory IP schemes are built around human actors.
Where a lawyer earns their fee
Because the law is unsettled, most of the value sits in getting the arrangements right before a problem appears.
A lawyer can advise on whether a particular workflow involves enough human contribution to support a copyright position. They can review AI platform terms for output rights, training use of your inputs, indemnities and liability caps. They can put IP assignment clauses and non-infringement warranties into contracts with contractors and developers, draft a workplace policy for acceptable AI use and confidentiality, and handle it when a rights holder or competitor comes knocking, including assessing infringement exposure and moral rights issues.
There is a separate set of obligations if your inputs contain personal information: feeding customer data into an AI tool can engage privacy law, particularly where the data is processed offshore. That is a distinct question from copyright, but it is part of the same pre-launch review.
The disputes that reach this area tend to be decided on evidence: what a human did, what the contract said, and what was published. Those are all things you can document cheaply now and reconstruct expensively later.
The evidence question to answer before you publish
Before AI-assisted material goes live, answer one question in writing: what did a human actually do, and what does the contract say about everything else? If the answer is "the tool did it and the platform terms say we own it", that is a risk position, not a legal position. If the answer is a documented trail of prompts, edits and selections, backed by contracts that assign rights and warrant non-infringement, you have a position you can defend.
Getting that in place early is a modest cost compared with unpicking it after a dispute. A short consultation can usually identify which of your workflows are exposed, and it is a conversation worth having before the next campaign goes live.