Generated images: what is unsettled, and the three rules that are not

Two separate legal questions, both open

People collapse these into one and then get confused. Keep them apart.

Was making the model lawful? Image generators were trained on very large collections of pictures, including copyrighted work, mostly without permission. Whether that training is lawful is being litigated in several countries at once, on different doctrines: fair use in the United States, text and data mining exceptions in the EU and UK, and other frameworks elsewhere. Cases have gone in different directions and appeals are outstanding. Nobody can tell you the answer, and anyone who does is selling something.

Do you own what comes out? Different question, different answers by country. The United States Copyright Office has taken the position that material generated purely by a machine lacks the human authorship copyright requires, while a work with sufficient human authorship in its selection and arrangement may be protected in respect of that contribution. The United Kingdom has a long-standing provision for computer-generated works with no human author. In March 2026 the government proposed removing it, while keeping protection for works made with AI assistance. No law has changed yet. Other jurisdictions differ again.

The practical consequence for a small business is narrow and worth stating plainly: an image you generated may not be yours to stop anyone else from using. For a logo or anything you intend to defend, that matters. For a decorative blog header, it usually does not.

None of this is legal advice, and the position where you trade may differ from all of it.

Three rules that are not unsettled

Whatever happens in the courts, these hold nearly everywhere.

Never depict a product you do not sell. A generated photograph of a dish you do not make, a room you do not have, a finish you do not stock. This is straightforwardly misleading advertising, and it is the fastest route to a complaint that sticks.

Never present a generated face as a real person. Not as a customer, not as staff, not as a testimonial. Even where nothing prohibits it specifically, it is a deception your customers can detect and will not forgive.

Never generate someone else’s brand, character or likeness. Trademark and personality rights do not become unclear merely because a model produced the image. A generated picture of a well-known character on your packaging is the same infringement it always was.

Labelling

Requirements are arriving rather than settled. The EU AI Act adds transparency duties for synthetic content as it phases in. Major platforms increasingly require disclosure of AI-generated media and add labels automatically when they detect it. Some images now carry provenance metadata under the C2PA standard, which platforms read.

Two practical points. Metadata often survives an upload, so an unlabelled image may be labelled for you. And in regulated trades (clinics, cosmetics, weight loss, financial services) a generated before-and-after image is the single most reliable way to attract a complaint that goes somewhere.

Where generated images are useful

Not photographs of your products. Real ones are better and you already have the products.

• Abstract or decorative backgrounds where nothing is being claimed.
• Illustrations and diagrams: how a process works, what a room layout looks like.
• Internal use: mock-ups, ideas to show a supplier, sketching a design before commissioning it.
• Social posts where the image is obviously an illustration rather than a depiction.

For anything that carries your identity permanently (the logo, the shopfront, the packaging) pay a designer once. It is a few hundred, it will be yours in a way a generated image may not be, and it will be better.

Free paths

Free tiers exist on every hosted generator with monthly caps. Stable Diffusion runs on your own machine through ComfyUI or Automatic1111 if you have a reasonable graphics card, with no per-image cost and nothing uploaded. Krita has a plug-in that puts local generation next to real drawing tools, which is a good combination for illustration work.

And keep the receipts: which tool, which prompt, what date. If provenance is ever questioned, that record is what you have, and it costs nothing to keep a text file.

What to tell a client who asks

Businesses that make things for other people (agencies, designers, printers, marketers) get asked whether AI was used and whether the client owns the result.

Answer both plainly. Say what was generated and what was drawn or photographed. Say that ownership of purely generated material is unsettled and differs by country, and that you are not giving legal advice. If the client needs to own and defend the asset, say so and price for human work.

Putting this in writing before the job protects both of you. The dispute that actually happens is not about copyright doctrine; it is a client discovering after delivery that something was generated when they had assumed otherwise.

Keep a human in the loop on the brief

Where generated imagery does work well, it works because a person made the decisions: what to show, what to reject, what to fix afterwards in a real editor.

That human contribution is also, in several jurisdictions, the thing that determines whether anything is protectable at all. Selecting, arranging, editing and combining are the activities that arguments about authorship turn on. Documenting them (a note of what you chose and why) costs a minute and is the only evidence you will have.