Learn how to license your likeness for AI generation, set enforceable limits and get paid. A step-by-step guide for creators — start here.
Photo via Unsplash
If you have ever searched how to license your likeness, you have probably found two kinds of answer: dense right-of-publicity law written for entertainment lawyers, or platform marketing that skips straight past the part where you keep control. Neither helps much when you are the person whose face it is.
This guide is the practical version. It walks through what a likeness licence actually contains, how to write limits a system will enforce rather than merely remember, what to check before you sign anything, and where creators most often get caught out. Everything here reflects how enforcement works on LikeWard, where a request that breaks your rules is refused before a model is ever called.
Key takeaways
- A likeness licence is permission on stated terms — not a sale. If it cannot be revoked, it is not a licence.
- Your limits should be structured data, not prose. Prose gets interpreted; data gets enforced.
- Ask three questions of any platform: what gets refused, what proof does each output carry, and how fast can I revoke?
- Four categories are refused platform-wide on LikeWard and no creator can switch them on.
- Compliance is now the commercial argument, not a cost centre — regulators, card networks and buyers all want the same artifact.
A likeness licence is a permission slip with edges. You are not handing over your face. You are saying: these uses are fine, those are not, here is how long that holds, and here is how I take it back.
The old version of this was a PDF. Someone at a brand read it once, filed it, and then a junior designer three years later made something nobody had cleared. The document was never wrong — it was just never consulted at the moment it mattered.
The useful version is a boundary set: the same decisions, written as structured data a system reads on every single request. You can see LikeWard's full vocabulary on the boundary types reference, and each category has its own explainer page.
Our marketplace desk describes the shift like this:
"The question stopped being 'what did she agree to?' and became 'what does the rule say?' Those sound similar. They are not. The first is an argument. The second is a lookup."
Start with the hard no. It is easier than starting with the yes, and it is the part you will care about at 2am.
On LikeWard this is the forbidden depictions list. You tick categories, and any request that lands in one is refused with that category named. Common picks:
Four categories are not yours to decide, because they are refused for everyone: content involving minors, non-consensual framing, deceptive news contexts, and criminal contexts. No boundary set can enable them. That is deliberate — it means you never have to argue about them with anyone, and it means the platform's floor does not move if a bigger creator asks nicely.
You edit the set and publish a new version. Boundary sets are append-only and versioned, so v3 does not quietly overwrite v2 — every past output still records which version authorised it. That is what makes an audit possible years later.
Forbidden depictions say what must never happen. Permitted contexts say where your likeness is welcome. Anything not on the list is refused by default, which is the right way round.
| Context | Typical creator | Notes |
|---|---|---|
| Portrait | Almost everyone | The safe baseline; headshot-style output. |
| Fashion editorial | Models, style creators | Pairs naturally with a brand exclusivity list. |
| Fitness | Athletes, coaches | Watch the overlap with health claims. |
| Gaming avatar | Streamers | Stylised by nature, so a low realism ceiling fits. |
| Brand endorsement | Anyone with sponsors | Add excluded competitor brands here. |
| Music performance | Artists | Often licensed to a label rather than personally. |
Table: the permitted contexts creators most often enable first, and what to watch for in each. Full definitions live on the boundary reference.
If you have signed an exclusivity deal — say, a drinks brand for eighteen months — list the competitors you are contractually barred from appearing with. The contract stops being a document somebody is supposed to have read and becomes a rule the system applies. That single feature is why brand campaign work is usually the easiest category to open up.
This one is underrated. A realism ceiling caps how photoreal any output can be, regardless of what else is permitted.
Three tiers, roughly: illustrative (clearly drawn), stylised (recognisable but obviously rendered), and photoreal. A request that asks for photoreal output from an illustrative-ceiling creator is refused with realism_exceeds_ceiling, and the refusal names both values so nobody has to guess.
Plenty of creators licence broadly on context and then pin the ceiling low. It is a good instinct. Most of the harm in this space comes from output that could be mistaken for a photograph, not from output that is obviously fan art. If you are unsure where to start, start at stylised — you can always raise it.
Here is the question that separates serious platforms from marketing: when I revoke, what happens to what already exists?
The weak answer is "we stop new generations". That is table stakes and it is not enough. Images already out in the world carry a provenance record, and if that record keeps saying authorised after you have withdrawn, your revocation is decorative.
The answer you want is that revocation flips the licence status on outstanding outputs too. On LikeWard, revoking marks every previously issued generation as revoked, the public provenance endpoint stops returning it as a live licence and withholds the output URL, and the manifest remains readable so distributed copies are still traceable back to the record that now says withdrawn.
That distinction is exactly what the US TAKE IT DOWN Act's 48-hour removal duty needs to be met in practice — we covered the mechanics in the TAKE IT DOWN Act compliance guide. Under the FTC-enforced regime, a platform cannot act on a report in two days if it has no idea which images are affected.
Immediately, and without asking anyone. If a platform's revocation route is a support ticket, treat that as the real answer to how much control you have.
Licensing pays in two shapes, and you want both explained before you agree to anything:
Get the split in writing, including whether it applies to both. Ask when payouts clear. Ask whether the platform can change the split unilaterally — if the answer is yes, that is a real term, not a technicality. LikeWard records splits per agency and shows them in the console; current pack pricing sits on the pricing page.
A word on expectations. Demand for licensed likeness is early. What is not early is the surrounding anxiety: a Google Ads Keyword Planner run we made in August 2026 shows "take it down act" averaging 9,900 US searches a month with a competition index of 3, and "age verification law" at 1,900. Nobody searches a statute for fun. Those are platform operators and managers trying to work out what they are now responsible for — and they are the people who end up buying licensed supply.
Safer than the alternative, which is unlicensed generation happening anyway with no record of your position at all.
Two pieces of law shape this directly. In the UK, the Online Safety Act 2023 puts duties on services that host user-generated content, with age-assurance obligations overseen by Ofcom. In the US, the TAKE IT DOWN Act (S.146) creates a notice-and-removal duty enforced by the Federal Trade Commission.
Neither says "do not license your likeness". Both say, in effect, that if imagery of a real person exists on a service, somebody needs to be able to prove what was consented to and act quickly when it was not. A published boundary set plus a signed manifest is that proof. A handshake is not.
Writing your limits as prose. "Nothing tasteless" cannot be enforced by anything. Pick categories.
Skipping the realism ceiling. Creators set contexts carefully and leave realism wide open. Then they are surprised by photoreal output in a context they did allow.
Not checking the refusal message. Ask to see one. A good refusal names the rule; a vague "request could not be completed" tells you the enforcement is probably a filter bolted on afterwards.
Assuming the API is looser than the website. On some platforms it is, which makes the website's rules theatre. On LikeWard the web app, the licensed-likeness API and the MCP server all run the same evaluation, and we have proven identical refusals from each.
Leaving the exclusivity list empty. If you have sponsors, fill it in on day one. The day it matters is the day you have forgotten.
Treating agency onboarding as paperwork. If you are managed, your agency configures this. Make sure you have actually read what they set. Agencies can run whole rosters from one console — that is convenient, and convenience is how nobody reads anything.
Ready to see what a live boundary set looks like before committing to one? Browse the boundary type reference, then read the boundary configuration guide for the field-by-field walkthrough. If you are managed, send your agency to the agency overview.