Why authorship is the sticking point
Copyright systems generally protect original works created by a human author. That requirement is old, uncontroversial and was written long before anything could produce a picture without a person making the creative choices in it.
A purely prompted image tests it directly. Somebody typed a description and a model produced pixels, and the question is whether describing a desired outcome amounts to authorship of the result, or whether it is closer to commissioning.
Several jurisdictions have reached the view that output produced without sufficient human creative control is not protectable. Others have not addressed it, and a few have specific provisions for computer-generated works that predate this technology entirely.
The practical consequence for a business is uncomfortable rather than catastrophic. An image you cannot own is also an image nobody else can own, which limits your ability to stop a competitor using the identical picture.
Where human contribution changes the answer
The more a person shapes the result, the stronger the argument for protection. This is the axis most legal commentary has settled on, even where the thresholds differ.
- Single prompt 0–20
- Iterated prompting 20–40
- Generated then edited 40–65
- Generation inside a human work 65–88
- Human work, generative touch-up 88–100
The right-hand end is the comfortable place to operate. A photograph you took, retouched with a generative tool, remains your photograph in every jurisdiction that has considered the question, because the authorship was never in doubt.
The left-hand end is where the difficulty concentrates, and it is exactly where most commercial use sits: a marketing image produced by describing what was wanted, with no underlying human work at all.
Why terms of service usually decide it first
Copyright is the interesting question and rarely the operative one. Every generator has terms, and those terms typically assign whatever rights exist, impose conditions, and reserve the ability to change both.
| Clause | What it usually says | Why it matters |
|---|---|---|
| Output assignment | Rights, if any, pass to the user | Cannot assign what does not exist |
| Commercial use | Permitted on paid tiers, restricted on free | The most common trip-up |
| Attribution | Sometimes required, sometimes not | Affects how you publish |
| Exclusivity | None, another user may get the same output | You cannot stop a lookalike |
| Indemnity | Offered by some vendors, capped | Matters if you are sued over training data |
The exclusivity row is the one that surprises people. Even where terms grant you everything the vendor has, the same prompt can produce a near-identical image for somebody else, and there is no mechanism to prevent it.
The indemnity row is the one procurement teams care about most. Several vendors now offer to defend commercial customers against claims arising from training data, and the presence, scope and cap of that promise is a real differentiator.
Where detection connects to the question
You cannot apply a rule you cannot enforce, and most organisations discover they have generated imagery in their published work long after it went out. Checking is how a policy becomes something other than a document.
There are three moments where it pays. Auditing your own back catalogue before somebody else does; checking supplier and agency deliverables against the contract you signed; and checking anything a competitor is asserting rights over.
The third is worth expanding on. Where somebody demands you stop using an image, a finding that the image is itself generated is directly relevant to whether they have anything to enforce, in the jurisdictions that have taken a position on it.
The training data question, separately
Ownership of the output is one dispute. Whether the model was entitled to learn from the images it was trained on is a different one, and it is the litigation that has attracted most attention.
These two questions travel together in conversation and apart in law. An image can be unprotectable as output and still expose its user to a claim arising from what the model was trained on, which is a genuinely awkward combination.
That is why vendor indemnities have become a procurement item rather than a footnote. An organisation publishing generated imagery at any scale is taking a position on a live legal question, and the question of who bears the cost if it goes badly is worth settling in advance.
For most users the practical answer is unglamorous: prefer vendors that offer an indemnity, keep a record of which tool produced which asset, and avoid prompting for named artists or recognisable styles, which is where the clearest claims arise.
Building a workable internal rule
Organisations that handle this well tend to have written three lines rather than a policy document, and the three lines are about where generated imagery may appear rather than about the technology.
First, no generated imagery in anything asserting a factual claim. News, case studies, testimonials, before-and-after photographs and anything depicting a real customer or a real event. This is the line that protects trust, and it is separate from the legal question.
Second, generated imagery is permitted for decoration and illustration, labelled where a reader might reasonably assume it was photographic. Backgrounds, abstract art, concept illustration: nobody is deceived and nobody objects.
Third, brand assets are photographed or commissioned. Anything that has to be defended against a copycat needs to be something you can own, and in several jurisdictions purely generated output is not.
Who is actually asking this question
The people who need an answer are rarely the ones reading case law. Three situations produce almost all of the real enquiries, and each has a practical resolution that does not require the law to settle first.
A designer asked to hand over rights in a deliverable. The honest answer is that they can assign whatever exists and cannot assign what does not, and a contract that says otherwise is describing something the law may not supply. Say which assets were generated and let the client decide.
A business building a brand identity on generated artwork. The exposure is not that somebody sues; it is that nobody can be stopped from using the identical mark. Anything that has to be defended should be commissioned or photographed.
A publisher who has received a demand to remove an image. Here the question is inverted and detection is directly relevant, because whether the complainant has anything to enforce may depend on how their image was made.