Can You Copyright AI Art? What the Rules Say

The governing document is the Copyright Office's statement of policy, published 16 March 2023 at 88 Federal Register 16190 and effective that day (US Copyright Office (2023), read 7 September 2026). Its reasoning is that copyright protects "only material that is the product of human creativity", and that a prompt does not supply that creativity.

Users do not exercise ultimate creative control over how such systems interpret prompts and generate material. Instead, these prompts function more like instructions to a commissioned artist.

— US Copyright Office, registration guidance, 16 March 2023

The Office has also published a second stage of analysis: Part 2 of its report, which "addresses the copyrightability of outputs created using generative AI", was published on 29 January 2025 (Copyright Office AI initiative (2026), read 7 September 2026). If you want the reasoning in full rather than the working rule, that is where it lives.

The Office describes a case-by-case test that turns on who executed the "traditional elements of authorship", the expression itself rather than the idea. Two routes are named explicitly in the guidance as capable of supporting a claim.

• Selection and arrangement: "a human may select or arrange AI-generated material in a sufficiently creative way that the resulting work as a whole constitutes an original work of authorship".
• Modification: "an artist may modify material originally generated by AI technology to such a degree that the modifications meet the standard for copyright protection".
• Neither route protects the generated material itself. Protection covers the human contribution, and nothing underneath it.

A worked example sits in the guidance itself. In February 2023 the Office concluded that a graphic novel combining human-authored text with images generated by an AI service was a copyrightable work, while the individual images inside it were not. That is the shape of most real claims: a protected whole containing unprotected parts.

So when people ask can you legally copyright AI art, the accurate answer is that you register a work, not a render. The more of the work you actually made, the more there is to register.

The guidance sets out an obligation and a mechanism, and both are quoted here because the wording matters.

1. 1

   Disclose

   "Applicants have a duty to disclose the inclusion of AI-generated content in a work submitted for registration and to provide a brief explanation of the human author's contributions to the work."

2. 2

   Exclude what is not yours

   "AI-generated content that is more than de minimis should be explicitly excluded from the application", using the Limitation of the Claim section under the Material Excluded heading.

3. 3

   Do not name the tool as an author

   The guidance says applicants should not list an AI technology, or the company that provided it, as an author or co-author simply because it was used.

4. 4

   Describe the human contribution briefly

   One or two plain sentences about what you selected, arranged or modified. This is the part examiners read.

5. 5

   Keep the file trail

   Drafts, prompts, edit history and dated exports. Nothing in the guidance requires them, and everything about a dispute makes you wish you had them.

The consequence of skipping step one is spelled out: where information essential to registrability "has been omitted entirely from the application or is questionable", the Office may move to cancel the registration, and a court may disregard a registration under section 411(b) of the Copyright Act where an applicant knowingly supplied inaccurate information.

Contract fills the gap. Where nothing is copyrightable nobody holds an exclusive right, so who owns AI art becomes a question about the tool's terms of service rather than about copyright: usually you hold a licence to use the outputs, sometimes with conditions attached.

• Read the terms for output rights, commercial use, exclusivity and any attribution requirement, per tool, not per industry.
• Assume non-exclusivity unless the terms say otherwise. Two users with similar prompts can receive similar outputs, and neither can stop the other.
• Check whether the terms differ by plan. Free tiers sometimes carry narrower commercial rights than paid ones.
• Client work needs the same clause twice: what you may deliver, and what the client may then do with it.

Provenance data is worth attaching even when ownership is unclear. Content Credentials, the C2PA-hosted standard, records how a file was made and edited, and the specification is public (C2PA specification (2026), Content Credentials (2026), both read 7 September 2026). A credential is not a title deed, but it is evidence of process, which is exactly what a contested authorship claim needs.

Yes, and the UK is the clearest contrast. Section 9(3) of the Copyright, Designs and Patents Act 1988 provides: "In the case of a literary, dramatic, musical or artistic work which is computer-generated, the author shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken" (legislation.gov.uk (2026), read 7 September 2026). That is a statutory answer to a question the US answers with a human-authorship test.

Sources for that table: EU AI Act guide (2026), CIPO (2026) and WIPO (2026). Two cautions. A transparency rule is not an ownership rule, and the absence of a published statement is not the same as a settled position. Neither of those is legal advice for your facts.

Usually yes, and that is a separate question from ownership. Selling depends on your licence from the tool and on the rules of wherever you are selling; copyright decides whether you can stop somebody else from copying it afterwards.

• The tool's terms decide whether you may sell the output at all.
• The marketplace decides whether it accepts generated work and how it must be labelled.
• Advertising rules decide how you may describe it: claims about a product or service have to be honest and substantiated (FTC (2026), read 7 September 2026).
• Copyright decides only whether you can enforce exclusivity, which for a bare generated image you likely cannot.

One practical note from our own workbench data (August 2026): the most common reason an external image job fails is the content-safety filter, with 49 blocks in 30 days. Commercial constraints start before the licence, at the point where a tool declines to produce something at all.

It is live, it is contested, and it is a different question from the one this page answers. Whether training on copyrighted material is lawful, and on what terms, is the subject of active litigation and policy work in several countries, and nothing here predicts an outcome.

What can be said from official sources: WIPO lists training data and copyright among the open questions it is convening discussion on, and the US Patent and Trademark Office maintains its own AI initiative covering the patent side of the same technology (USPTO (2026), read 7 September 2026). The Copyright Office signalled in its 2023 guidance that it would seek public input on how the law should treat the use of copyrighted works in AI training, and its notice of inquiry followed later that year (Copyright Office NewsNet (2023)).

For a working creator the practical exposure is not the training question at all. It is producing something too close to an identifiable existing work, which is an ordinary infringement question with an ordinary answer: do not do it, and do not prompt for it by name.

Five habits, none of them expensive.

1. 1

   Record your process

   Keep prompts, references, intermediate files and the dates. This is the evidence for the human-contribution explanation you will have to write.

2. 2

   Disclose when you file

   The duty is explicit, and the penalty for ignoring it lands on the registration you were trying to protect.

3. 3

   Read the tool terms once, properly

   Output rights, commercial use, plan differences. Ten minutes, then keep a copy of the version you agreed to.

4. 4

   Do not assume exclusivity

   Plan your business on the parts you actually made: the edit, the brand, the series, the distribution.

5. 5

   Attach provenance where you can

   Credentials that record how a file was made cost nothing and answer the question everyone asks second.

So, can you copyright AI art? The human parts, yes, on the record and with the generated material disclaimed. Everything else you protect by contract, by process and by making more of the work yourself. If you generate in our image generator, the settings panel and the AI photo editor both leave a trail worth keeping, and every run records which model produced it, which for most image jobs is GPT Image 2.