Copyright of AI-generated works is the body of rules and case law that decides whether output produced by an artificial intelligence system can be protected by copyright, and if so, who owns it. In the field of Copyright & AI, the dominant answer in the United States and the European Union is that protection requires human authorship: material generated autonomously by a model is not protected, while human creative contributions such as selection, arrangement and substantive editing can be.
The question is distinct from whether training a model on protected works is lawful, which is governed by the text and data mining exception, fair use and licensing. Here the focus is on the output side: who, if anyone, holds rights in what the system produces.
Why human authorship is the decisive test
Copyright protects original expression created by a person, so the decisive question for any AI output is how much of its expressive content reflects human creative choices. Most copyright statutes never defined an author explicitly, but courts and registries have long read the concept as referring to a natural person, and generative AI has forced that assumption into the open.
In the United States, the Copyright Office applies the human authorship requirement and refuses registration for material produced without sufficient human control. The courts confirmed this in Thaler v. Perlmutter, where an image generated autonomously by a system called the Creativity Machine was denied protection; the District of Columbia Circuit affirmed that refusal in March 2025. In the European Union, protection depends on the work being the author’s own intellectual creation, a standard from Court of Justice case law that requires free and creative choices reflecting the personality of a human author.
The practical consequence is that purely machine-generated content typically falls outside copyright and can be copied by others, unless contract law, trade secrets or other rights restrict that use.
How prompts, editing and selection affect protection
Human input is assessed by its degree of control over the expressive result, not by the effort invested. The Copyright Office report on copyrightability, published in January 2025, concluded that prompts alone are generally insufficient, because the same prompt can yield many different outputs and the model, not the user, determines the expressive details.
The following table summarizes how common forms of human involvement are usually treated under the current US approach, which European registries and courts broadly echo.
| Human input | Typical treatment | What may be protected |
|---|---|---|
| Simple text prompt | Usually insufficient | Nothing in the raw output |
| Iterative prompting | Usually insufficient | Nothing, even after many rounds |
| Human-made input image | Protected where it remains perceptible | The human-created portion |
| Selection and arrangement | Can qualify as a compilation | The arrangement, not each item |
| Substantive modification | Can qualify | The human changes |
| Region-level editing tools | Assessed case by case | Depends on the control exercised |
The Zarya of the Dawn decision of February 2023 illustrates the split: the Office registered the text written by the author and the selection and arrangement of images in a comic book, but excluded the individual images generated with Midjourney.
Approaches in major jurisdictions
Jurisdictions agree on the human-centered principle in outline but differ in the details, and a few have special rules for computer-generated works. The table compares the positions most often relevant to international publishers and developers.
| Jurisdiction | Legal basis | Position on AI output |
|---|---|---|
| United States | Copyright Act and Office practice | No protection without human authorship |
| European Union | Originality standard from case law | Requires the author’s own intellectual creation |
| United Kingdom | Section 9(3) of the 1988 Act | Computer-generated works protected for 50 years |
| China | Court decisions since 2023 | Some outputs protected where prompts show creative input |
| Japan | Copyright Act and agency guidance | Protection only for creative human contribution |
The UK rule assigns authorship of a computer-generated work to the person who made the arrangements necessary for its creation, but it predates generative AI, sits uneasily with the originality requirement and has been under government review. The Beijing Internet Court decision of November 2023, which protected an image produced through detailed prompting and parameter changes, shows that outcomes may diverge across borders for the same output.
Ownership, disclosure and contractual terms
Where protection exists, ownership follows the ordinary rules: the human author owns the protected contribution, or the employer under work-for-hire or employment rules. The AI provider does not become an author by supplying the tool, although its terms of service may assign, license or disclaim rights in outputs as a matter of contract.
Registration practice adds disclosure obligations. Since its March 2023 guidance, the US Copyright Office requires applicants to identify AI-generated material that is more than de minimis and to claim only the human-authored parts. Failing to disclose can lead to cancellation of the registration and weakens later enforcement.
| Party | Basis of any rights | Typical limits |
|---|---|---|
| User or creator | Human creative contribution | Only the human-authored elements |
| Employer | Employment or commissioning rules | Same scope as the employee’s contribution |
| AI provider | Contract and terms of service | No authorship; rights mainly contractual |
| Rights holders of training data | Infringement claims if output copies their works | Must show substantial similarity |
Risks, misconceptions and practical safeguards
A frequent error is to assume that paying for a tool or writing a detailed prompt automatically creates ownership. Unprotected output gives no exclusivity, so competitors may reuse it, and a lack of protection does not exclude infringement risk if the output reproduces a protected work memorized from training data.
- Document the human workflow, including drafts, edits and selection decisions, to prove creative contribution.
- Separate human and machine material in registrations and internal records.
- Review provider terms for output ownership, indemnities and usage restrictions.
- Screen outputs for similarity to known works before commercial release.
- Do not rely on moral rights for unprotected content, because they attach only to human authors.
The law remains unsettled and fast-moving: courts have decided relatively few cases, legislative proposals are pending in several countries, and the line between assistive and generative use will keep shifting as tools gain finer controls.
Put simply
Copyright rewards people for creative work. If a machine produces a picture or a text on its own, most countries say no person created it. So nobody owns it in the copyright sense. Anyone may copy it.
Things change when a person shapes the result. Writing parts, choosing and arranging pieces or heavily editing them all count. Then the human part can be protected. Typing a request into a chatbot is like giving instructions to an artist. The instructions alone do not make you the author.
In practice
A small publishing house prepares an illustrated children’s book. The author writes the full story herself, and a designer generates about 400 candidate illustrations with an image model, selects 24, repaints faces and backgrounds by hand and arranges them across the pages.
When registering the book in the United States, the publisher discloses the AI-generated elements and claims the text, the hand-made modifications and the selection and arrangement. The registration is granted on that basis. When a merchandise seller later copies one illustration onto mugs, the publisher can enforce rights only to the extent the copied image contains human-authored changes, which is why the designer’s saved edit history becomes the key piece of evidence.