Skip to content
AI & legal work

Who owns AI-generated content — and what you can actually enforce

Every generative tool of any size tells you the output is yours. That sentence resolves one question — the vendor will not turn round and claim your file — and quietly leaves the more important one open. Copyright is not granted by a contract. It arises, or fails to arise, by operation of law, and in the United States it requires a human author. Where that leaves you depends entirely on how much of the finished work you actually made.

7 min readPublished How we write these

The short version

  • Purely machine-generated output is not copyrightable in the United States. The Copyright Office requires a human author, and prompts alone do not make you one.
  • "You own the output" in a vendor's terms is a contractual promise that the vendor will not claim the file. It is not evidence that a copyright exists in it.
  • What is protectable is your own contribution: expressive material you put in, substantial edits you make afterwards, and the selection and arrangement of generated elements.
  • AI material that is more than de minimis must be disclaimed when you register. Registration is also the gate to an infringement suit, so an unprotected logo cannot be enforced against a copyist.

Two questions that get answered as one

Ask who owns AI-generated content and the answer arrives from the vendor: the terms of service say the output is yours. That is true, and it is narrower than it sounds. It settles who, as between you and the company that ran the model, may claim the file. It does not establish that anything ownable was produced.

A clause transferring "all right, title and interest" in the output transfers whatever the vendor had. Where the output is purely machine-generated, that is nothing. The clause functions as a covenant — the vendor will not compete with you over the file, and will not license it out from under you. It is not a title deed, and no drafting can make it one.

On 29 January 2025 the US Copyright Office published the second part of its report on copyright and artificial intelligence, dealing with copyrightability. Its conclusion was that existing law is adequate and no new right is needed, because the existing test already answers the question: copyright protects human authorship, and the extent of protection tracks the extent of the human contribution. Prompts, however long and however many, were held insufficient on their own — they convey the idea but do not control the way it is expressed.

The courts have gone the same way on the outer edge. In Thaler v. Perlmutter the D.C. Circuit held in March 2025 that human authorship is required as a matter of statutory law, and the Supreme Court declined to review that decision in March 2026. Thaler is a narrow fact pattern — the applicant named his machine as the sole author and disclaimed any creative contribution of his own — so it fixes one end of the range rather than the middle.

How much of you is in the file

Nothing protectable
The arrangement, not the images
Ordinary copyright, no asterisk

Prompt only

Prompt, then curation

Output materially reworked

Your work, AI-assisted

The two middle bands are where every real dispute sits, and neither the Office nor a court will tell you in advance which side of the line a particular file falls on.

How much reworking is enough is genuinely unsettled. The artist Jason Allen entered more than six hundred prompts into Midjourney, used its variation and upscaling tools, and cleaned up the result in Photoshop; the Office's Review Board still refused registration, and his challenge to that refusal is before a federal court in Colorado. Anyone telling you where the line falls is guessing.

What is protectable in a work that contains AI material

The registration granted for the graphic novel Zarya of the Dawn is the clearest worked example. The Office registered the text the author wrote and the selection, coordination and arrangement of the images. It excluded the Midjourney images themselves. One work, two answers, on the same certificate.

  • Expressive input. Where you feed in your own drawing, photograph or draft and the model alters it, the parts of your original that remain perceptible in the output stay protected, on the same reasoning that applies to any derivative work.
  • Selection and arrangement. Choosing which generated elements to keep, in what order, on what page, is authorship in the compilation — the same protection a photo editor has in a picture book.
  • Substantial modification. Edits that would themselves be original if you had made them to anyone's work are protected as your contribution. Colour-correcting is not that. Redrawing is.

The terms of service are a promise about the vendor

It is worth crossing the two facts, because most people treat one of them as decisive and it never is.

What the terms say against what you contributed

Human authorship in the file

What the vendor's terms say

Silent, or a broad licence back

"You own the output"

Prompts only

Nobody owns it

No copyright arises, and the vendor has not even promised to stay out of your way. Public domain in all but name.

A promise, not a right

The vendor will not claim the file. Neither can you claim it against a copyist, because there is no copyright to assert.

Substantially your expression

Yours regardless

Your expression, your copyright, whatever the terms omit. Check separately what licence back the vendor took.

Ownership and clear title

A copyright exists and the vendor has waived what it might have argued. The only cell worth building a brand on.

Only the bottom-right cell gives you something to enforce. The top-right is where people believe they are, and it is the cell that reads best in a pitch deck.

Graphic design contract template

Sets out the deliverables, the rights that transfer, and a disclosure clause covering whether generative tools were used in producing them — the question a client should be asking before, not after, the logo goes on the packaging.

Open

Registering a work with AI material means declaring it

Since March 2023 the Copyright Office has required applicants to disclose AI-generated material that is more than de minimis, and to disclaim it in the application rather than quietly leaving it in. A model, or the company that made it, must never be named as an author or co-author. Where an application has already been filed without the disclosure, supplementary registration is the mechanism for fixing it.

Where the disclosure goes on the application

An application covering a work that contains AI material

The excluded material is not a defect in the application. Leaving it out of the exclusion is.

The incentive to leave the disclosure off is obvious and the exposure is worse than the omission. A registration obtained on an application that concealed material facts can be cancelled, and it hands any defendant a ready-made attack on the certificate before the merits are reached.

An uncopyrightable logo cannot be enforced

This is the part that turns an abstract doctrine into a bill. Suppose a competitor copies your generated brand illustration pixel for pixel. For a US work, section 411(a) of the Copyright Act makes registration a prerequisite to bringing an infringement action at all. If the Office refuses registration because the image has no human author, you can still sue — the statute lets you proceed once you have served the Register — but you are now litigating the authorship question first, at your own cost, with a hostile agency determination already on the file.

Trademark is a separate route and it does not have this problem. Trademark law is about source identification, not authorship, so a generated logo can function and register as a mark provided it is distinctive and used in commerce. What that protects is the mark's job of telling customers who made the goods — a competitor using a confusingly similar sign in a related trade. It does not stop someone reusing the same artwork on an unrelated product, which is the gap copyright would ordinarily have filled. Our guides on registering a trademark and on which right covers what set out the difference.

What to write into the contract

When you commission creative work in 2026

  • Ask directly whether generative tools were used, and for what. "We use AI in our workflow" is not an answer; which deliverables, at which stage, is.
  • Require the supplier to identify any material they cannot warrant as human-authored, rather than warranting the whole file and hoping.
  • Do not accept an assignment clause as proof of ownership. An IP assignment moves whatever exists; where nothing exists it moves nothing, silently.
  • For anything that will carry a brand, ask for source files and a record of the human editing steps. That record is the evidence of authorship if it is ever tested.
  • Where the deliverable is licensed rather than assigned, check the content licence actually restricts the supplier from relicensing the same generated asset to a competitor — copyright will not do it for you.
  • Budget for the possibility that a key visual is unprotectable, and decide in advance whether that matters for this particular use.

Where this lands

The doctrine is stable even though the boundary is not. Machines do not author works; humans do, and copyright covers what the human brought. What has changed is that the question is now worth asking about ordinary commercial output — a logo, a product illustration, a set of marketing images — where nobody used to ask it, because nobody used to produce those things without a person drawing them. The uncomfortable answer for a lot of files generated in the last three years is that they belong to no one, are freely copyable, and came with a contract that said otherwise. That is not a reason to stop using the tools. It is a reason to know which of your assets you could actually defend, and to make sure the ones that matter have a person in them. Whether the same logic applies to work a freelancer produced for you is a related question with a different answer.

General information, not legal advice. This guide explains how these documents and rules generally work. Law varies by jurisdiction and changes, and none of it is applied to your circumstances here. For anything consequential, consult a licensed attorney where you are.

Frequently asked

Can I copyright an image I made with Midjourney or DALL-E?

Not the image as generated. The US Copyright Office treats purely machine-generated output as lacking a human author, and prompting — however detailed or iterative — has not been accepted as authorship on its own. What you can register is your own contribution: text you wrote, substantial edits you made to the output, or the selection and arrangement of generated elements within a larger work.

The terms of service say I own the output. Is that wrong?

It is correct and it is answering a different question. The clause allocates rights as between you and the vendor, so the company will not claim your file or license it elsewhere. It cannot create a copyright that does not exist. Where the output has no human author, the vendor has assigned you something with no exclusive rights attached to it.

How much editing makes AI output copyrightable?

There is no threshold anyone can state, and the Copyright Office decides case by case. The registered example to reason from is Zarya of the Dawn, where written text and the arrangement of images were protected but the images themselves were not. Edits that would count as original authorship if applied to someone else's work will generally count here; routine cropping and colour correction will not.

Do I have to tell the Copyright Office that I used AI?

Yes, where the AI-generated material is more than de minimis. Since March 2023 the Office has required applicants to disclose it and disclaim it in the application, and never to list a model or its developer as an author. If you have already filed without disclosing, supplementary registration corrects the record — which is a far better position than defending a concealment.

Can someone legally copy my AI-generated logo?

If the artwork has no human author, there is no copyright in it to infringe, so copying the image itself is not a copyright wrong. Trademark law can still help: a logo used in commerce as a source identifier can be registered as a mark and enforced against confusingly similar use in a related trade. That is narrower protection than copyright and it takes registration and use to establish.

Do the whole thing on your phone

Draft it, check it for risk, rewrite the clauses you do not like, sign it and send it — without opening a laptop.

  • 136 templates across 12 categories
  • AI review in plain English
  • Free every month — 3 documents, 2 reviews
Download on theApp Store
Free to download · no account

iPhone, iPad, Mac & Vision Pro · iOS 15.6+ · 76.1 MB
Premium from $1.99/week