Two open questions and one you can answer today

A live disagreement, not a settled rule

This lesson does not tell you what the law is, because in most of the world it is unresolved, actively litigated, and moving. What it can do is set out the shape of the disagreement clearly enough that you can make your own commercial decisions and know which questions to ask a lawyer.

Anyone who tells you confidently that AI-generated images are definitely fine to sell, or definitely infringing, is describing their position rather than the law.

Question one: was the training lawful?

Generative models are trained on very large collections of existing work, much of it copyrighted and much of it gathered without individual permission.

The argument that it is lawful holds that training is an analytical process which extracts statistical relationships rather than storing and republishing works, and that this falls within existing exceptions (fair use in some systems, text and data mining exceptions in others) much as indexing the web has been treated.

The argument that it is not holds that copying at scale to build a commercial product is reproduction regardless of what happens next, that exceptions were written for research rather than for commercial model training, and that the output competes directly with the work it learned from.

Where it stands: multiple significant cases are in progress in several countries, with divergent early results and appeals pending. Some jurisdictions have introduced text and data mining exceptions with opt-out mechanisms for rights-holders. Others have not. Several have proposed transparency obligations requiring disclosure of training data. This is unsettled, and settlements between parties have resolved some disputes without establishing any general rule.

Question two: who owns the output?

A distinct question, with more visible divergence.

Some copyright offices and courts have taken the position that material generated without sufficient human authorship is not protected by copyright and falls outside the system: meaning nobody owns it and anyone may use it. In such systems, human contribution to the final work can attract protection, but the purely generated elements may not.

Other jurisdictions have provisions for computer-generated works that assign authorship to the person who made the arrangements necessary for the creation. These predate modern generative models and their application to them is untested.

The commercial consequence for a client is direct: an image that is not protected by copyright cannot be exclusively theirs, and a competitor may be free to use it. This matters enormously for a logo or a brand asset and much less for a background texture.

Question three: what do the provider’s terms say?

The question you can actually answer today, and the one most people skip.

Every generative tool has commercial terms, and they differ. Read the current version for the specific tool, and check: whether commercial use is permitted at all or only on a paid tier; what the provider claims about output ownership; whether any indemnity is offered against infringement claims, and what conditions attach to it; whether your inputs are used for further training; and whether outputs must be disclosed as AI-generated.

These terms change, sometimes materially. Save a dated copy of the ones in force when you delivered the project.

What a working creative can actually do

Five practices that are defensible under any of the plausible legal outcomes.

Keep human authorship in anything that must be owned. Where a client needs exclusive rights (a logo, a brand mark, a signature illustration) make sure there is substantial, documented human creative work: your drawing, your composition, your decisions. Keep the evidence. This protects the client under every version of the output question.

Disclose, and put it in the agreement. Say what you used and where. A client discovering it later is a trust failure regardless of the legal position, and this belongs in the written agreement rather than in conversation.

Do not promise exclusivity over purely generated elements. You cannot honestly guarantee that what nobody may own is exclusively theirs.

Do not prompt in the style of a living artist for commercial work. Whatever the copyright position on style, doing so invites claims under passing off, unfair competition and personality rights in several jurisdictions, and it is ethically poor practice in a small industry.

Keep records. Which tool, which version, which prompts, what you did afterwards, the terms in force, the date. If a question arises in three years, this file is the entire answer.

Talking to a client about it

Be plain and do not pretend to certainty you do not have:

“I used a generative tool for the background textures and drew the mark myself. The law on who owns purely generated material is unsettled and differs by country, so I would not want to promise that those textures are exclusively yours. The mark is original human work and that part is straightforward. If exclusivity on everything matters to you, we should produce the textures directly and I will quote for that.”

Most clients respond well to this, because it is more information than they usually get, and it lets them make the decision that is theirs to make.

What not to do

Do not give legal advice. Do not tell a client that AI-generated work is safe, or that it is protected, or that training was lawful. Say what you did, say what is unsettled, and recommend they take advice if the stakes justify it.

Do not assume a rule from one country applies in another. The divergence here is unusually large, and the jurisdiction that matters may be the client’s, not yours.

The limitation

The honest limitation of this lesson is its subject. Positions are being taken by courts, legislatures and copyright offices continuously, and some of what is described here will have moved by the time you read it. Any specific case outcome cited today may be reversed on appeal, and any statutory position may be amended.

That is not a reason to avoid the tools or to pretend the question is simple. It is a reason to build the habits that survive whichever way it resolves: human authorship where ownership matters, disclosure in writing, no promises you cannot keep, and a record of what you did: all of which are good practice independent of any court.