AI Image Licensing: Generation Is Not Permission

This guide helps you decide whether an AI-generated image can go into a commercial project, and what records to keep before it does. One caveat up front: it covers practice and risk, not legal advice. Ownership and liability depend on platform terms and local law.

Generation is not a licence

That the model can produce an image is a technical fact. Whether that image can go into a client deliverable is a question of licensing and compliance. The two get conflated constantly, which produces two familiar accidents: using free-tier output on paid work, and shipping an image carrying someone else's trademark or a celebrity face.

Split the problem into three parts and it becomes manageable: does the platform license you to use it commercially, does the content infringe anything, and does it pass content review. All three must clear.

Three things to check first

Check these before you generate, in this order.

1. Platform terms: is commercial use allowed, is your output used for training, how is ownership worded, is attribution required.
2. Source material: whose reference image is it, and does the prompt contain a trademark, a living artist's name or a real person's likeness.
3. Generation record: prompt, parameters, timestamp, account and model version. If a dispute arises, this is the only evidence chain you have.

The third is the most neglected and the cheapest to fix. Records cannot be reconstructed after the fact.

Why platform terms cannot be generalised

Rules differ widely and change with each revision. The differences cluster around four points: whether free and paid tiers have different commercial rights, whether output may be used for training, how ownership is described, and whether AI generation must be labelled.

Midjourney, for example, grants commercial use on paid plans from Basic upward, offers no free tier since the trial was withdrawn in 2023, and excludes trial output from commercial use. That is its position, not a universal rule.

Domestic platforms generally tie commercial terms to the membership agreement, and free tiers often carry extra restrictions. So the right move is not to memorise a conclusion but to build a habit: read the terms when a project starts, read them again before delivery, and archive screenshots of the commercial use, ownership and liability sections. If the terms are revised, your dated screenshot marks the boundary.

Three high-risk prompt patterns

• Third-party trademarks: logos, packaging, storefront design, signature colour combinations. If the output reads as a known brand, you already have exposure.
• Living artists' names used for imitation: "in the style of" a named artist is the classic red flag, especially when the prompt asks for their signature technique.
• Real people's likeness: celebrities, clients, colleagues. Do not use an unauthorised face, including generating a lookalike from a photo.

Better alternatives: describe style in photographic language such as hard high-contrast light, low saturation, soft focus or wide-angle distortion, rather than naming a person. Use synthetic faces or licensed models. Have the client supply official brand assets for compositing instead of asking the model to invent them.

Where reference images get risky

Style influence and direct copying are different things, and the line is drawn by platform terms and local law. This article does not offer a legal conclusion, but three practical rules apply immediately:

• Do not run someone else's finished image through image-to-image and call it yours, especially commercial or photographic work.
• Do not reproduce another image's composition, character design or signature elements. Push the reference until only the mood remains.
• Keep provenance: shot by you, licensed, or explicitly permitted by the platform. Untraceable references should not enter a project.

If an image leans heavily on a reference, keep the provenance note and licence before commercial use. When a client asks, documentation beats explanation.

Content review is stricter in China

Publishing or advertising on Chinese platforms means stricter review than most overseas tools assume. Political sensitivity, sexual content, violence, likeness of real people, maps and national symbols are all common rejection triggers. Medical, financial and education advertising carries additional limits on promised outcomes.

Before generating, ask whether the image would survive review. Portrait work is the usual trouble spot: a highly realistic synthetic face may require labelling or be refused outright. When in doubt, change the concept rather than gamble on passing.

A pre-launch checklist

• Read the platform terms, confirm your tier allows commercial use, and archive a screenshot.
• Verify sources: references are licensed, and the prompt has no trademark, artist name or real likeness.
• Keep generation records: prompt, parameters, timestamp, model version and account.
• Remove client or third-party logos, storefronts and packaging from the frame.
• Use synthetic or licensed faces, and sign a model release for any real person.
• Set type in post rather than generating it, to avoid garbled characters.
• Keep the original export and layered files, and disclose which assets are AI-generated.
• Write AI asset clauses into the contract: scope of licence, sublicensing, and who carries the risk if platform terms change.

The last item is the one most often skipped. Without a clause, a dispute falls back to defaults and everyone argues. With one, both sides know where the boundary sits.

Three internal rules for a team

Individuals can rely on judgement; teams need process. Three rules are enough:

1. Allowlist: only platforms and tiers confirmed for commercial use. Tools outside the list do not touch client work.
2. Paper trail: one asset folder per project containing prompt, parameters, original file and terms screenshot.
3. Second review: before anything ships, someone else checks the high-risk list, focusing on trademarks, faces and text.

The conclusion is one sentence: being able to generate something is not the same as being allowed to sell it. Before commercial use, do three things: read the platform terms, verify the sources, keep the records. Turning those three into a process is more reliable than memorising any single rule.

To build up the tooling picture first, read the AI image tools introduction. To standardise your production flow, see the AI image workflow guide. For further compliance and asset references, browse the resource library.