The short answer is that it depends entirely on which tool you used and under which plan, and that the answer changed several times in the last two years. The longer answer is worth understanding before you build a business on output you do not own.
This is not legal advice. It is the practical map of what the licences actually say, and the specific places where people get burned.
The Three Questions That Decide It
1. Can you use it commercially at all? Some models are released for research only. Others permit commercial use but exclude specific industries. Qwen-Image, for example, permits commercial use under its licence; other open-weight models carry restrictions that rule out commercial work entirely while still calling themselves open.
2. Who owns the output? For most hosted tools, you own what you generate subject to their terms, on a paid plan. On free tiers, several services retain broad rights to use your output — read the free-plan terms, not the marketing page.
3. Does the training data create downstream risk? This is the unresolved one. If a model was trained on copyrighted work without permission, the question of whether its output infringes is unsettled, and different jurisdictions are landing in different places.
Where People Actually Get Burned
Assuming "open" means "free to use commercially". It frequently does not. Open-weight means you can download the weights; the licence attached to those weights is a separate document with separate rules. Some restrict commercial use above a revenue threshold, some require attribution, some exclude entire categories of use.
Assuming a paid plan transfers ownership. Paid plans usually grant broad commercial rights to you, but the mechanism differs by provider — some grant a licence, some assign rights, some are vague. For most commercial work the practical effect is the same. For anything you intend to license onward or register, the distinction matters.
Ignoring voice and likeness. Voice cloning and generating images resembling a real person are the two areas with clear, immediate legal exposure that has nothing to do with AI copyright law. Consent and personality rights apply regardless of how the output was made.
Not keeping records. If you ever need to demonstrate what you generated, when, and under what terms, you will want export logs and plan receipts. Almost nobody keeps these, and they are the only evidence you will have.
By Output Type
Images. The cleanest position is a commercial hosted plan from a major provider, or an open-weight model whose licence explicitly permits commercial use. The murkiest is a model with a research-only or non-commercial licence — which is a real category, not a hypothetical one.
Music. Generated music is the most legally unsettled category, because the outputs are closer to existing copyrighted works and the industry litigates aggressively. Suno and Udio both grant commercial rights on paid plans; what that protects you from is contested, not settled. Lyria sits in the same unsettled space.
Voice. The model output is rarely the issue; the cloned voice is. Clone your own voice or one with documented written consent, and keep the consent document.
Text. The lowest risk category, with the obvious caveat that generated text can still reproduce or closely paraphrase source material. Run anything client-facing through a plagiarism check before delivery.
A Practical Checklist
Before you ship AI-generated work commercially:
- Confirm the tool's licence permits commercial use — and read the actual licence, not the summary
- Confirm your plan tier grants commercial rights (free tiers often do not)
- If using open weights, check revenue thresholds and attribution requirements
- For voice, obtain and store written consent before cloning anyone
- For music, understand you are operating in unsettled territory and avoid exclusive reliance
- Keep export logs, prompts and receipts
- For client work, disclose AI use in the contract — several industries now require it
What Is Actually Settled
Being straight about the limits of this advice:
- Settled: purely AI-generated output without human authorship is not copyrightable in the US. You cannot register it, which means you cannot sue over it in the way you could over authored work.
- Settled: contracts and licences between you and the tool provider are enforceable, and they govern most of what you will actually do.
- Not settled: whether model training on copyrighted work makes output infringing. Active litigation, different outcomes by jurisdiction, no reliable answer yet.
The practical implication: your protection comes from the tool's licence and your contract with your client, not from copyright in the output itself.
Browse our Image category for tools and their terms.
Frequently Asked Questions
Can I copyright AI-generated images?
In the US, no — not without substantial human authorship. This matters because it means you cannot register the work and therefore cannot bring an infringement claim over it the way you could over authored work.
Does a paid subscription mean I can sell what I make?
Usually, but not always, and the scope varies. Paid plans from major providers generally grant commercial rights; free tiers frequently retain rights to your output. Check your specific plan terms.
Are open-weight models safer commercially?
Often more predictable, because the licence is a document you can read, but not automatically more permissive. Some open licences restrict commercial use entirely or above a revenue threshold.
What is the biggest actual legal risk?
Voice and likeness rights. Cloning someones voice or generating their image without consent creates exposure under personality rights and, in some places, criminal law — entirely separate from AI copyright questions.
Should I disclose AI use to clients?
Yes, increasingly as a contractual requirement rather than a courtesy. Several industries now mandate disclosure, and being the one who mentioned it first is a much better position than being the one caught.
