Who Owns AI-Generated Art? A Plain-English Answer

This isn't legal advice. We're a wall art company, not a law firm. But because this question comes up almost every week, here's our best understanding of the current state of things in plain English, with links to read further.

The short version

As of early 2026, in the United States: the raw output of a text-to-image AI model — generated purely from a prompt, with no human editing beyond writing the prompt — is generally not copyrightable. You cannot register a pure prompt-only image with the US Copyright Office and enforce a copyright on it against third parties. However, images that include significant human authorship (meaningful editing, compositing, painting-over, art direction with reference images, etc.) can be partially copyrighted, with the AI-generated portions excluded from protection.

Most countries have landed in roughly the same place, though with variations. The UK is slightly more permissive; the EU is slightly stricter; China has gone in several directions simultaneously.

What this means if you buy a print from us

• You own the physical print. Nobody can take it away from you. You can hang it, give it as a gift, resell it as an object.
• You do not own a copyright on the underlying image. You can't stop someone else from generating a similar image, because the image isn't copyrightable in the first place (and also because you didn't create it — we directed it).
• For most customers, this is fine. You bought a print for your wall, not a licensing empire.

What changed in 2025

The US Copyright Office issued updated guidance clarifying that human-directed AI workflows — where a person makes meaningful creative choices, iterates, edits, and exercises art direction — can result in partially copyrightable works. The AI-generated raw output still isn't protected, but the human layer on top can be. This is a big deal for studios like ours: it means the art direction we put into pieces is the protectable layer, and the raw generation isn't.

There's also been movement on training-data lawsuits — specifically, the Copyright Office and several courts have leaned toward treating training on copyrighted works as a fact-specific question, with outcomes depending on how transformative the use is. This is still evolving, and we watch it closely.

What to ask a seller

If you care about this, ask three questions before buying:

1. "Was the model trained on licensed or opt-in data?" If yes, you're on cleaner ground.
2. "What rights am I getting with the print?" (Usually: the physical object, non-exclusive personal-use rights to the image, no license to reproduce commercially.)
3. "What happens if someone finds the image offensive or files a takedown?" A serious seller has an answer; a lazy one doesn't.

What we do

Our model is trained on a mix of licensed corpora and public-domain works, with no opt-outs required because we didn't scrape anything we shouldn't have. We attach a brief human-art-direction document to every order showing the creative decisions our designers made on top of the raw generation — both for transparency and, frankly, to preserve whatever protectable authorship exists in the final piece. We don't claim copyright we don't have. We don't allow reselling of our images. And if someone raises a good-faith concern about a specific piece, we deal with it directly rather than hiding behind a terms-of-service clause.

Where to read further

The US Copyright Office's AI guidance documents from 2023–2025 are surprisingly readable. The Verge, 404 Media, and Platformer have the best ongoing coverage of the specific court cases. Cory Doctorow's newsletter is opinionated but informed. Your actual lawyer, if you have one, is better than any of us on the specifics of your situation.