Four questions, routinely mashed into one
People ask "can I use AI images commercially" as though it were a single question. It is four, they have different answers, and only one of them is genuinely unsettled.
- Was the model trained lawfully?
- Are you licensed to use this model?
- Do you own the output?
- Does the output infringe somebody else?
None of what follows is legal advice. It is the shape of the problem. A lawyer in your own country is the person who answers it for your case.
One: was the model trained lawfully
Live litigation, several countries, no settled answer. Getty sued Stability in the US and the UK. Artists have brought class actions in the US. Publishers and image libraries have brought proceedings in Europe. Some cases have narrowed, some have settled quietly, none has produced a clean rule.
You will not resolve this and you do not have to. What you can do is decide how much exposure to it you want to carry, which leads to the tools that offer indemnity, below.
Two: are you licensed to use the model
This one has a definite answer and you can look it up in two minutes. It is lesson one's licence trap.
- Open weights carry a licence. Apache 2.0 releases such as FLUX.1 [schnell] and Qwen-Image permit commercial use. FLUX.1 [dev] does not. Stability's recent community licence is free below a revenue threshold and paid above it.
- Hosted tools carry terms of service, and free tiers frequently restrict commercial use where paid tiers do not.
These positions were accurate in 2025 and they change. Read the model card or terms page on the day the job starts.
Three: do you own the output
Genuinely unsettled, and different by country. The short version:
- The United States requires human authorship. Prompts alone do not make you an author. Your own expressive contributions — a drawing you fed in, your edits, your arrangement — can be protected.
- The United Kingdom and a group of jurisdictions with similar provisions, including India, deem the author of a computer-generated work to be the person who made the arrangements for its creation.
- China's Beijing Internet Court reached the opposite conclusion to the US on similar facts, finding protection where the user's choices showed personal intellectual investment.
- The European Union has no harmonised answer and its member states are working it through.
The same image can be unprotectable in one country and protected in another. Making things with AI sets out the cases in detail.
The practical consequences: a prompt is not property, "you own your outputs" in a tool's terms is a promise not to claim it rather than a right the company can create, and for a logo you should route around the problem entirely — trademark protects a mark through use and registration however it was drawn.
Four: does it infringe somebody else
This is the risk that actually bites, and it is the one people ask about least. A recognisable cartoon character, a real person's face, a trade dress, a style so specific it names a living artist. That exposure exists whether a human or a model drew it.
Indemnification is what Adobe, Getty and Shutterstock offer on business plans: if a claim arrives about the training data behind their generator, they will stand behind it. Read the exclusions. It covers what the model learned from. It does not cover you prompting a famous mouse.
Labelling: moving fast, in one direction
- China has required both a visible label and a metadata label on synthetic content since September 2025, with a duty on platforms to check.
- India amended its IT Rules for synthetically generated information and prescribed *visibility* — a label covering a defined portion of the frame and the opening portion of audio, plus a platform duty to verify user declarations.
- The European Union's AI Act requires machine-readable marking of generated output and disclosure of deepfakes; the timetable has been amended more than once, so check the current position.
- The United States has no general federal labelling law, a 2025 statute on non-consensual intimate imagery, and a patchwork of state election rules.
The technical layer — C2PA Content Credentials, and invisible watermarks such as Google's SynthID — is real and fragile. A screenshot destroys it. Many re-uploads strip it. It is worth attaching and it is not a substitute for a label a person can see.
The line for telling a client
Tell them, in writing, whenever a reasonable viewer could take the image as a record of something real: a person, a place, a product's actual appearance, a customer, an event. Tell them also when they will license the asset onward, because the ownership question becomes theirs and they need to know what they are selling.
One line in the delivery note does it:
Assets 3, 5 and 7 are AI-generated. They do not depict real people or real events. Generated with [tool] under [plan]; source files and settings supplied.
That sentence has never lost anybody a job. Discovering it later has.
Today: write that line into your own delivery template, before you need it.
Before you move on