Who Owns What a Model Produces

Your vendor’s terms almost certainly say you own the output. That resolves the position between you and the vendor and settles nothing about copyright, because a contract can only transfer rights that exist. Whether any copyright exists in machine output depends on the jurisdiction, and the answers differ more than in almost any other area covered here.

Two different questions, routinely conflated

Question one is contractual. As between you and the model provider, who has the rights? Every major provider’s terms assign whatever rights exist in the output to the customer, and some disclaim any rights of their own. This is real and useful. It stops your vendor asserting rights against you and it lets you build a business on the output.

Question two is proprietary. Does a copyright exist at all, that you could assert against a third party who copies your output? This is a matter of law and no contract touches it. If no copyright subsists, your vendor has assigned you an empty set and a competitor may copy your generated marketing images with impunity.

Almost every business question that matters — can I stop somebody copying this, can I license it exclusively, does it count as an asset — depends on question two.

United States: human authorship required

The position is now reasonably clear and it is restrictive.

• Thaler v. Perlmutter (D.C. Circuit, March 2025) affirmed that the Copyright Act requires a human author, so a work listing a machine as its sole author cannot be registered. This settles the extreme case.
• The Copyright Office’s copyrightability report(January 2025) is the practical guide. Its conclusions: existing law is adequate and no new legislation is needed; a work is protected to the extent of its human authorship; prompts alone, however elaborate, do not give the user sufficient control over the expressive elements to make them the author of the output; but human selection, arrangement, and modification of AI-generated material can be protected, as can human-authored material that a model assists with.
• Registration practice follows from that. In the Zarya of the Dawn decision (2023) the Office registered a comic book’s text and the selection and arrangement of its images, but not the Midjourney-generated images themselves. Applicants must disclose AI-generated material that is more than de minimis and disclaim it.

The practical consequence for a business: a fully generated image has no US copyright. A layout you designed incorporating generated elements has copyright in the layout. A generated draft you rewrote substantially has copyright in your contribution. In each case the protected scope is narrower than the artefact, and if you ever need to enforce, you will have to identify which parts are yours.

United Kingdom: the computer-generated works section

The UK is one of a small number of countries with an express provision. Section 9(3) of the Copyright, Designs and Patents Act 1988 says that for a literary, dramatic, musical or artistic work which is computer-generated, the author is taken to be the person by whom the arrangements necessary for the creation of the work are undertaken. Section 178 defines computer-generated as generated by computer in circumstances such that there is no human author of the work. The term is shorter than for a human-authored work — fifty years from creation — and moral rights do not apply.

On its face that is a clean answer: somebody owns the output, and it is whoever made the arrangements. Two problems have never been resolved.

• Who made the arrangements? The user who wrote the prompt, the developer of the application, or the company that trained the model? There is no authority resolving this for generative models, and the answer is commercially significant.
• Does the provision survive the originality requirement? UK originality has been shaped by CJEU case law requiring a work to be the author’s own intellectual creation, reflecting free and creative choices. A work with no human author cannot satisfy that on any natural reading. Whether section 9(3) operates as an exception or is simply inconsistent has not been tested.

The government consulted on removing or amending section 9(3) in 2021 and 2022 and decided to leave it alone, citing lack of evidence of harm either way. The 2024–25 copyright and AI consultation raised it again. It remains in force, untested and quietly uncertain.

European Union: no equivalent provision

There is no computer-generated works provision in EU law. Copyright subsists in a work that is its author’s own intellectual creation — the standard established in Infopaq (C-5/08, 2009) and developed in Painer (C-145/10, 2011), which requires free and creative choices stamping the work with a personal touch.

Machine output with no human creative choices does not meet that standard, and there is no fallback. The analysis is therefore closer to the US one than to the UK one, with the same practical result: what is protected is the human contribution, and identifying it is your problem if you ever enforce.

China: courts have found copyright

China is the significant counterexample and it is usually left out of comparisons.

In November 2023 the Beijing Internet Court held, in a case concerning an image generated with Stable Diffusion, that the image was copyrightable and that the user was its author. The reasoning turned on the user’s intellectual investment: the prompts written, the parameters set, the iterative selection and refinement across many generations were held to constitute original intellectual achievement reflecting the user’s personal expression. That is a materially lower bar than the US Copyright Office applies to the same conduct.

Earlier, in 2019, a Shenzhen court held in the Dreamwriter case that an automatically generated financial news article was a work, protected as a work created by a legal person under the organisation and responsibility of the company.

Two consequences. The same generated image may be protected in China and unprotected in the US, which makes “is this protected?” a question with no jurisdiction-free answer. And Chinese courts have also found output infringement — the Guangzhou Internet Court in 2024 found a platform liable where its service generated images substantially similar to a well-known character — so the permissiveness runs in one direction only.

Patents: settled, and settled the other way

The inventorship question was litigated in parallel and resolved consistently across jurisdictions: an inventor must be a natural person. The UK Supreme Court so held in Thaler v Comptroller-General in December 2023; the US Court of Appeals for the Federal Circuit reached the same conclusion in Thaler v. Vidal in 2022; the European Patent Office’s Legal Board of Appeal did likewise.

None of that prevents patenting an invention made with AI assistance. It requires a human inventor to be named, which in practice means the human contribution to the conception has to be real and documented. Patent offices in several jurisdictions have issued guidance on AI-assisted inventions saying essentially that.

What protects the work when copyright does not

The absence of copyright is less catastrophic than it first appears, because copyright is rarely the only protection and often not the relevant one.

What to do in practice

1. Decide which assets need to be assertable. Most generated content does not. Marketing copy that will be stale in a month does not need copyright; a character design your product is built around does. Spend the effort only where enforcement is a real prospect.
2. For the assets that matter, keep a record of the human contribution. Drafts, edits, selection notes, the versions you rejected. In the US this is what supports a registration; in the EU it is what supports subsistence. Retrospective reconstruction is not the same thing.
3. Disclose AI material when registering in the US.Failing to disclose more than de minimis AI-generated content risks the registration itself. The Office has cancelled registrations over this.
4. Read the vendor’s output clause for what it does not say. It says the vendor claims nothing. It does not say the output is original, that it does not reproduce training data, or that you can stop others copying it. Those are three separate risks and only the second is usually covered, by indemnity, on conditions. See the clauses that matter.
5. Do not assume similar outputs are your infringement problem. If your generated asset resembles an existing work, the exposure is a claim against you, and it is governed by ordinary substantial-similarity analysis. Whether your own output is protectable is a separate question from whether it infringes somebody else’s.