What You Can Actually Copyright in AI-Assisted Art

Key Takeaways

• A prompt alone doesn't make you the author. In the US, purely AI-generated images can't be copyrighted, however detailed the prompt.

• Your human contributions can be protected. Your own drawings or photos that show up in the result, your edits to AI output, and your creative selection and arrangement of AI material all count.

• Disclose the AI parts when you register. The US Copyright Office requires you to identify AI-generated material and claim only what you made.

• Keep your process. Sketches, layered files, edit history and screen recordings are what prove the human part.

• Platform terms don't create copyright. When a tool says you own its output, it means the company won't claim it, not that the law protects it.

The short answer

Copyright protects human creativity. In the US, that means the parts of an AI-assisted artwork that came from you can be protected, and the parts the AI generated on its own can't. Most real work is a mix, so the practical question is which parts are yours and whether you can show it.

This guide covers US law in detail, because that's where the clearest guidance and most of the cases are, then looks briefly at the UK, the EU and China. It's a general explainer, not legal advice. For anything with money on the line, talk to a copyright lawyer.

What the US Copyright Office says

The US Copyright Office set out its position in a January 2025 report on copyrightability. Its central conclusion is that copyright requires a human author, and that prompts alone don't give a person enough control over the output to count as its author. A prompt works more like instructions to a commissioned artist: the person who writes them doesn't own the resulting drawing.

The report lists three ways human contribution can still be protected:

Your own expressive inputs. If you feed in your own drawing, painting or photo and it remains recognizable in the result, your original work is protected within it.

Selection and arrangement. Choosing and arranging AI-generated elements in a creative way, such as a comic layout or a collage, can be protected as a compilation, even when the individual images aren't.

Modifications. Edits you make to AI output, such as painting over it, compositing or substantially reworking it, can be protected to the extent they're your own creative work.

The Office decides each case on its facts and said it doesn't think new legislation is needed.

The cases that drew the lines

Thaler v. Perlmutter. Stephen Thaler tried to register an image with his AI system listed as the sole author. The D.C. Circuit ruled in March 2025 that the Copyright Act requires a human author, and the Supreme Court declined to hear the case in March 2026. The ruling settles that a machine can't be the author. It deliberately left open how much human involvement is enough.

Zarya of the Dawn. In 2023 the Office narrowed the registration for Kristina Kashtanova's comic book. Her text and her selection and arrangement of the text and images were protected; the individual Midjourney images weren't.

A Single Piece of American Cheese. In January 2025 the Office registered an image made by Invoke's Kent Keirsey from about 35 separately inpainted regions. It was refused at first, then registered after he submitted a timelapse of his process. The registration covers his selection and arrangement of the regions, not the AI-generated pieces themselves.

Allen v. Perlmutter. Jason Allen, whose Midjourney image Théâtre D'opéra Spatial won a Colorado art prize in 2022, is suing after the Office refused to register it unless he disclaimed the AI material. He argues that more than 600 prompts and his Photoshop work make him the author; the government says they don't. Briefing finished in February 2026 and the case is still pending. It could become the first court ruling on where the line falls for prompt-heavy work.

What that means for common workflows

• Text-to-image from a prompt, used as is: Not protectable in the US, however long or detailed the prompt.

• Many generations, picking the best one: Choosing a single output isn't enough on its own. Selecting and arranging many outputs into a larger work can be.

• Your sketch or photo turned into a finished image: Your original sketch or photo is protected where it's still visible in the result; the parts the AI added aren't.

• AI output you paint over, composite or heavily edit: Your edits can be protected, and the more of the final image is your work, the stronger your claim.

• Inpainting region by region under close direction: The American Cheese registration shows the arrangement can be protected, with good process records.

• A comic, book or game that uses AI images: Your writing, story, layout and arrangement are protected even if individual images aren't.

How to register AI-assisted work

The Copyright Office's 2023 registration guidance still applies. When you apply, you must disclose AI-generated material that's more than trivial, describe what you contributed, and exclude the AI material from your claim using the application's fields for excluded material. If you registered AI-assisted work before that guidance, you're expected to correct the record with a supplementary registration. Failing to disclose can cost you the registration.

Disclosure doesn't weaken your claim to the parts you made. It's what gets those parts registered.

What to keep as proof

• Your inputs: sketches, reference photos and any original artwork you fed into the tool.

• Your process: prompts, settings, every iteration and the outputs you rejected.

• Your edits: layered source files, such as PSDs, and edit history.

• A recording: a screen recording or timelapse of the work in progress. Invoke's registration turned on one, and a Chinese court ruled against an artist in 2025 partly because he couldn't show his process.

Platform terms aren't copyright

Most AI tools say you own what you create. OpenAI's terms, for example, assign you all its right, title and interest in output, "if any." Midjourney's terms say you own your images to the fullest extent the law allows. That language means the company won't claim your images or stop you using them commercially. It can't give you a copyright the law doesn't recognize, so purely AI-generated output may still be free for anyone to copy.

Outside the US

United Kingdom. The UK is unusual: section 9(3) of its copyright law gives "computer-generated works" with no human author a 50-year copyright, owned by whoever made the arrangements for the work to be created. The government's March 2026 report on copyright and AI proposes removing that provision, but it's a proposal, and no bill has been introduced. AI-assisted work that shows human creativity stays protected under the normal rules either way.

European Union. EU law protects works that are the "author's own intellectual creation," which courts read as requiring a human making free, creative choices. A Prague court ruled in 2023 that a DALL-E image wasn't protected. The EU AI Act's transparency rules, in force from August 2026, are about labeling AI content, not copyright.

China. Chinese courts have gone both ways. The Beijing Internet Court granted copyright to a Stable Diffusion image in 2023 because the user's prompts, settings and selection showed intellectual input, and a Changshu court protected an edited AI image in 2025. But a Zhangjiagang court denied protection in 2025 for an image made from a simple prompt whose creator couldn't show his process.

The bottom line

• Do the creative work you want to own. Your drawing, your edits and your arrangement are what copyright protects.

• Record it as you go. Process evidence is what separates a registrable work from an unprotected one.

• Be honest on the application. Disclose the AI parts and claim the rest.

• Watch Allen v. Perlmutter. It's the case most likely to tell artists where the line sits.