You Can Sell What the AI Made. You May Not Own It.

Two questions get tangled together every time this comes up, and separating them makes the whole subject easier. Can I sell this? Do I own this? They have different answers, they are decided by different bodies, and the second one matters far less to most sellers than the noise around it suggests.

A seller producing generated designs for prints, patterns, covers or digital downloads needs to get three things right: the disclosure, the human contribution, and the records. Everything else is commentary, except the production file itself, where resolution and colour decide whether the printed item matches the mockup.

Do you own what the model produced?

Not the raw output, and the reasoning is now official rather than speculative. In January 2025 the US Copyright Office published the second part of its report on artificial intelligence, and the analysis of its human authorship conclusions is unambiguous on the central point: fully AI generated outputs produced in response to a prompt lack human authorship, and therefore cannot be copyrighted.

Two further findings matter more to a working seller than the headline.

Entering a prompt is not authorship, however elaborate the prompt is. And choosing one output from a batch of twenty is not authorship either, because selecting a single output is not itself a creative act. Those two together demolish the most common workaround people describe, which is to generate a hundred variations and curate.

What does count is everything you add on top. The Office found that human contributions are sufficient for authorship where the human retains creative control over expressive elements, that modifying an output gives protection to the perceptible human expression in the result, and that protection extends to the selection, coordination and arrangement of human and machine material even where the machine parts remain unprotected on their own.

This is the same boundary we traced for two adjacent cases, and the three articles agree because the law does. A generated logo cannot be copyrighted but can still function as a trademark, which we worked through in the piece on the AI logo you cannot copyright but can register as a mark. And in music the protectable part is the human authorship layered on top, covered in what you actually own in an AI assisted track. Different products, one principle.

Why does ownership matter less than sellers think?

Because copyright is a right to stop other people copying you, and most sellers of generated goods were never going to exercise it anyway. Ask what you would actually do if you found your design on another shop tomorrow. Enforcing a copyright means identifying the seller, sending a formal notice and being willing to follow it, and the cost of that exceeds the value of a single design many times over.

What sellers really lose when a product is unprotectable is narrower and worth naming precisely. You cannot licence it exclusively, because you cannot grant what you do not hold. You cannot credibly promise a buyer exclusivity, which is a real commercial product in some markets. You cannot register it, which forecloses some enforcement routes in jurisdictions where registration is a precondition. And you cannot assign it in a sale of the business, which matters if you ever sell the shop.

If none of those four apply to how you trade, unprotectable output is a smaller problem than the discussion around it implies, and the disclosure question is the one deserving your attention. If any of them do apply, particularly exclusivity, then the human contribution is not a compliance detail. It is the product.

There is a second asymmetry worth understanding. Your inability to own the output does not give you any freedom regarding other people's rights. A generation that reproduces a protected character, a distinctive style associated with a living artist, or a trademarked logo is a problem for you whether or not you own the result. The two questions are independent, and only the second one produces claims against you.

What do marketplaces actually require?

Disclosure, in most cases, and the definition of what triggers it is stricter than sellers assume. The clearest published version comes from Amazon's self publishing arm, and the test it applies generalises well beyond books.

According to a detailed guide to the KDP disclosure rules, Amazon requires you to declare AI generated text, images including cover art, and translations, and it counts as generated even if you edited it substantially afterwards. Content you wrote that a model then helped you refine, proofread or error check is assisted, and needs no declaration. The line is drawn on direction of creation: whoever produced the first draft determines the category.

Two details in that policy are worth knowing before you decide how to feel about it. The declaration is not shown to buyers, so it is a compliance record rather than a label on your product page. And failing to declare risks removal of the title, account level strikes and potentially termination across your whole catalogue, while declaring accurately carries no stated penalty. That asymmetry is the entire decision.

Handmade and craft marketplaces work on a second axis as well as disclosure, which is the claim you make about how the item came to exist. A seller guide to the current Etsy standards reports that the platform expects Designed by rather than Made by when a model generated the core design, requires the disclosure to sit visibly in the listing rather than buried, and prohibits selling bundles of prompts as products in themselves. The same guide describes enforcement clustering around high volume uploading, hidden disclosures, and prompting for branded styles.

Does the EU require you to label AI output?

It requires marking, and the obligation falls mostly on the tool rather than on you. Article 50 of the AI Act sets transparency duties, and the text of the article puts the machine readable marking of synthetic audio, image, video and text on the providers of the generative systems, who must ensure the marking is technically feasible and reliable.

The duties on deployers, meaning the people using such a system, are narrower than the coverage implies. They cover informing people who are interacting with an AI system where that is not obvious, disclosing deepfake content, and disclosing artificially generated text published on matters of public interest. There are exemptions for assistive editing that does not substantially alter content, and for artistic, creative, satirical or fictional works, where only the existence of the generated material need be disclosed. The obligations apply from 2 August 2026.

For a seller of generated designs the practical reading is that the AI Act is not the reason to disclose. The marketplace rules are, and they arrived first and bite harder. The Act matters more for the marketing around the product than for the product itself, which is territory we covered in what you can and cannot claim about AI in your own copy.

Doing the disclosure without undermining the sale

The wording is the part sellers agonise over, and it is easier than it looks once you stop treating it as an apology. A disclosure that reads like a confession invites a buyer to interpret it as one. A disclosure that reads like a description of your process reads like craft.

Say what the tool did and what you did, in that order, in one sentence, in the listing description where a scrolling buyer will meet it. Something in the shape of: the base image was generated with an AI tool, then hand recoloured, composed and prepared for print by me. That satisfies a visible disclosure requirement and simultaneously tells the buyer where the human work is, which is the thing they are deciding whether to pay for.

Two mistakes to avoid. Do not put it only in a policies page or an FAQ tab, because a requirement to disclose in the listing means the listing. And do not use language that implies more human involvement than there was, because an overstated disclosure is worse than a plain one: it converts a compliance question into a misrepresentation question, which is a different and more serious category.

Put it in the template rather than deciding per item. A seller listing at any volume will eventually forget, and the forgotten one is the listing that gets reported.

What should you keep, and for how long?

Keep enough to reconstruct the human contribution, because that is the only evidence that distinguishes a product you can defend from one you cannot. Nobody asks for this until somebody does, and by then the files are gone.

Save the working file with layers or history intact, not only the flattened export. Save the original generation alongside the finished piece, so the difference between them is visible at a glance. Note the tool and the date. If you drew, photographed or wrote any component yourself, keep that source file separately, because it is the strongest single piece of evidence you will ever have.

Keep the licence terms of the tool as they stood when you generated, not as they stand today. Generative tool terms change, and what matters for work you produced in March is what the terms said in March. A dated screenshot or a saved copy costs nothing.

One more record is worth keeping and almost nobody does: the prompt itself, saved with the output. It is not evidence of authorship, as the Copyright Office made clear, but it is excellent evidence of intent when a platform asks whether you were deliberately imitating a protected style. A prompt that names no artist and no brand is a short answer to a question that would otherwise take paragraphs.

Two years is a sensible floor for a small seller, and longer for anything still on sale. The trigger for needing these records is almost never a lawsuit. It is a platform appeal, where you have a short window to demonstrate that a listing is what you said it was.

Can I say a generated item is handmade?

No, and this is where sellers get into real trouble rather than theoretical trouble. Handmade is a factual claim about production, it is what buyers on craft platforms are paying a premium for, and a false version of it is both a policy breach and a straightforward misrepresentation to a consumer. Designed by is accurate and available.

Will disclosing hurt my sales?

Less than sellers fear, and the evidence for the fear is thin. Buyers of digital downloads and print files are frequently indifferent, buyers of craft items are not, and the honest response to that difference is to sell into the market that wants what you actually make. Disclosure also removes the single largest downside risk, which is losing the account that carries your whole catalogue.

What about training data and copyright claims?

That litigation concerns the model makers rather than you in most scenarios, but the indemnity question is worth checking. Some tools offer a contractual indemnity for commercial use of their output and some explicitly do not, and it is stated in the terms rather than the marketing page. Read it before you build a product line on a tool, not afterwards.

Does any of this apply to my product photography?

Yes, and it is the case sellers overlook, because the image is not the product. A generated or heavily altered image of a physical item you sell is a representation of that item, so the ordinary rules against misleading depiction apply on top of any AI specific policy. Showing a colour, texture or scale the buyer will not receive is a returns problem before it is a compliance one.

The version of this that is worth building

Sellers who do well with generated products are rarely the ones producing the most listings. They are the ones who use generation for the part that was slow and keep doing the part that was theirs, which turns out to be the same thing the Copyright Office describes when it talks about creative control over expressive elements.

That gives you a working test with no legal training required. If a competitor could type your product description into any generator and get something close to your item, you have a volume business competing on price against people with lower costs. If they could not, because the value is in a composition, a house style, a set of components you made or a curation nobody else could repeat, you have something to own and something to disclose without embarrassment.

Build the disclosure into your listing template so it is never a decision you make item by item, keep the working files, and check the tool's terms once a quarter. If you are setting up the storefront that carries all of this, our AI store builder generates the listing structure on code you own outright, which means the disclosure field and the product template are yours to change rather than a platform's to withdraw.

Sell what the model helped you make. Claim only the part you actually did. Those two sentences cover almost every question in this article, and they are also just an accurate description of what you are selling.