The most common legal question about AI images is whether you own what you generated, and the answer surprises most people.
The confusion comes from conflating two separate matters: permission to use and copyright protection. The first is contractual, the second statutory, and they answer differently.
This piece separates them and covers the training-data risk that is discussed least but matters most commercially.
The two questions
First: may I use this image? The answer is in the terms of service. Using a tool means accepting its terms, and those terms define where your output may be used. This varies by tool and by plan.
Second: is this image mine? The answer is in copyright law. Copyright requires a work to be the product of human creativity, and writing a prompt is not, in most legal systems, treated as sufficient creative contribution.
Those diverging answers produce an odd result: you can use an image commercially and entirely lawfully while holding no exclusive rights over it at all.
Why ownership usually fails
Copyright statutes generally require a work to bear the personal stamp of a human author — the phrasing varies by jurisdiction but the requirement does not.
In fully AI-generated output the chain of human choices is limited to prompt writing, and the prevailing reading is that this does not create authorship.
The practical consequence: you can use the image, but the ground for stopping someone else using the same one is weak.
The picture changes as human contribution grows. Where you have substantially edited the output, combined several outputs or redrawn the result as a vector, that human contribution can be protected. This is why, for a logo intended for trademark registration, having a designer produce the final drawing is sounder.
The third issue: training data
Discussed least, but the most concrete commercial risk. Most image models were trained on material scraped from the internet with unclear rights.
That makes the chance your output resembles someone's copyrighted work non-zero. The risk is not theoretical: a t-shirt design or ad visual resembling a recognisable character can trigger an infringement notice and be pulled from a platform.
Three ways to lower it. Do not name brands, characters or living artists in prompts. Reverse image search before commercial use. For corporate work, prefer a model trained on licensed data — Adobe Firefly is in that category and offers indemnification on enterprise plans.
A checklist before commercial use
These six checks take minutes and prevent most of the problems that become expensive later.
- Read whether your tool AND your plan permit commercial use
- If you run an open-weight model, check the licence of the exact version
- Reverse image search the result; change it if it resembles a known work
- Keep brand, character and living artist names out of your prompts
- If a person's likeness is involved, obtain explicit consent
- If a trademark is planned, have a human produce the final drawing and document it
Frequently asked questions
Do I own the image I generated?
The right to use it is usually yours; copyright most likely belongs to nobody. In many countries, works without human creative input cannot be copyrighted. You can use the image but will struggle to claim exclusive rights over it.
Can I sell images made on a free tier?
It depends on the tool and must be checked. Some restrict commercial use on the free tier; others keep free-tier images public. On client work the second point often matters more.
Do I have to disclose that I used AI?
In most cases there is no legal requirement, but some platforms expect AI content to be labelled, and for news or anything making a factual claim disclosure is ethically necessary. This page is general information, not legal advice.