Clear Rights Before Publishing a Generated Image

"Can we publish this?" is not one question. It is four, they are independent, and passing one says nothing about the others. A team that only checks the first one — the service's terms — regularly ships images that are perfectly licensed and still a problem.

This lesson is a checklist, not legal advice. Rules differ substantially by jurisdiction and are actively changing; treat it as the set of questions to bring to whoever advises you.

Four independent gates

The licence gate asks what the service's terms allow. Read for three specifics rather than skimming: whether commercial use is permitted at your plan tier, whether the terms differ for free and paid accounts, and what happens to your rights if you stop paying. Model weights have licences too, and open-weight models sometimes carry acceptable-use restrictions that travel with outputs.

The depiction gate asks what is in the image, and this is the one that catches people. A perfectly licensed image can still depict a recognisable person, a trademarked logo, protected trade dress, a copyrighted character, or a distinctive building. The service's licence grants you nothing here, because these rights belong to third parties who were never party to your terms.

The authorship gate asks whether you can claim rights in the output. In several jurisdictions, purely AI-generated images attract limited or no copyright protection, which matters if the asset is meant to be a defensible brand element. A logo you cannot protect is a weak logo.

The context gate asks what the image will be taken to assert. The same picture is decorative in one placement and a factual claim in another, and the second one carries obligations the first does not.

The depictions that create risk

Worth memorising, because they recur:

• Recognisable real people. Publicity and personality rights apply even where the image was generated rather than photographed, and "it is not really them" is not a defence when the point was that it looks like them.
• Logos and trade dress. Models reproduce these from training data readily. Remove them or composite the licensed asset.
• Named artists' styles. Prompting "in the style of" a living artist is a legal grey area in most places and a reputational problem in nearly all of them. Describe the visual qualities you want instead — this also produces better and more distinctive results.
• Landmarks and buildings. Some jurisdictions restrict commercial use of images of certain structures.
• Anything implying a person endorsed something. This is advertising regulation, not copyright, and it is enforced.

The practical instruction: run a depiction pass on every image before it ships, looking specifically for faces, marks, and text. It takes fifteen seconds and it catches most of what matters.

Disclosure and context

Whether a synthetic image needs a label depends on what it is doing in the layout. A generated abstract background on a landing page asserts nothing. A generated photograph of a product asserts that the product looks like that. A generated image of people using your service asserts that people use your service.

Editorial, news, and documentary contexts are stricter than marketing, and several jurisdictions and platforms now require disclosure of synthetic media in specific circumstances. Some advertising regimes require disclosure whenever a demonstration is simulated.

The workable rule: if a reasonable viewer would treat the image as evidence of something, label it or do not use it. If it is decoration, context usually carries it — check your platform's rules anyway, since they are changing quickly.

Record provenance while you still remember it

Six months later, someone will ask where an image came from. Whoever is asked will not be you.

Keep, next to the asset: the model and version, the service and account tier, the prompt and seed, whether any input image was used and what its rights were, the licence conclusion you reached, and the date you reached it. The input-image row is the one that matters most, because an image-to-image output inherits the rights problems of its input — a generation seeded with a stock photo is a derivative of that stock photo, whatever the model did to it.

Where the tooling supports content credentials or similar provenance metadata, keep them attached, and know that most social platforms strip metadata on upload. Provenance you rely on should live in your asset system, not in the file's headers.

Practice

Take five images you have already published and complete the record retroactively. The ones you cannot reconstruct are your process gap, and they are usually the image-to-image outputs where nobody noted the input.

Then write a one-page rule for your team covering: which model and tier is approved, what depictions are prohibited outright, when disclosure is required, and what must be recorded at generation time. One page that people read beats a policy that nobody opens.

Common mistakes

Assuming the service's licence settles it. It answers one of four questions.

Prompting a living artist's name. Grey legally, bad reputationally, and unnecessary — describe the qualities.

Losing the input image's provenance. The most common gap, and the one that turns a clean output into a derivative work of unknown status.

Treating a free-tier generation as commercially cleared. Tiers frequently differ on exactly this point.