How-To Guide
How do you license creator content for AI edits?
Separate editing, synthetic performance and training rights, then write the ten fields an AI clause needs before a creator grant is signed.
License AI editing as its own grant, not as part of ordinary editing permission. Name the three possible uses separately: AI help with production on footage the creator actually made, a synthetic performance that puts new words or scenes in the creator's face or voice, and model training or fine-tuning on the creator's files. Say which one the brand is buying, what it may generate, where and for how long, who can touch the files, and what gets deleted at the end. Pay for it as a separate line, because a broad 'create derivative works' clause does not tell anyone what the brand intends. Then keep four questions apart: whether the contract grants the right, whether the creator approved this exact output, whether the platform requires a label, and whether the advertising is truthful. Passing one does not pass the others.
- Ordinary editing permission covers length, format, captions and cleanup, not synthetic performance or model training.
- The AI clause must name the inputs, outputs, training answer, replica answer, approval, term, vendors and deletion duty.
- Keep the contract right, creator approval, platform label and advertising truth as four separate gates.
Start with the permission, not the tool
A broad AI clause rarely arrives labelled as one. It arrives inside a production sentence, usually something like the right to modify, adapt, create derivative works from, or use artificial intelligence with the creator's name, image, likeness, voice, content and raw files. That one sentence can describe a caption tool, a translated ad, a synthetic endorsement, or a model trained for years. The words do not tell you which. Creators Agency puts the first rule plainly: do not treat a broad AI clause as ordinary editing permission.
So do not argue about the word AI in the abstract. Ask what outcome the brand is trying to reach. Which file goes into the tool, and what comes out? Can the output make the creator appear to say or do something they never recorded? Does the tool keep the data or use it to train another model? Who can access the result, and where will it run? Each answer turns a vague worry into a commercial term. The same page draws the line that matters: routine production assistance and synthetic performance are different permissions. Captions, transcript cleanup, noise reduction and colour correction help make the deliverable the creator already agreed to. Generating a new scene, line or endorsement makes a performance the creator never gave.
This is why the first negotiation is about scope, not price. If the brand cannot explain what it wants to do, the permission is too broad to price responsibly, and a number attached to it is guesswork. Ask for the workflow in plain language, then write that workflow into the clause.
One AI request, three permissions. Original diagram built from the Creators Agency, Snippet and Lathrop GPM clause checklists read for this page on November 18, 2026.
Make it work for your brand
Find it. Remake it. Post it. Superdirector turns the viral formats in your niche into your brand's carousels, X posts and video scripts, and keeps the calendar full.
Separate editing, synthetic performance, and model training
Three permissions hide inside one AI request, and each has a different value and a different exit problem. Editing help changes how real footage looks or reads. Synthetic performance creates words, scenes or a voice the creator did not perform. Model training uses their files to build a capability that outlives the campaign. Consent to the first is not consent to the other two, and a contract that bundles them gives away more than the buyer usually needs.
| Permission | What it lets the brand do | What the creator should ask for |
|---|---|---|
| AI-assisted editing of real content | Captions, translation of captions, cleanup, retouching, colour, cropping and platform exports on footage the creator actually recorded | A narrow production grant, with accuracy review where language or claims can shift, and no change to the creator's meaning |
| Synthetic performance or replica | Generate new lines, scenes, dubbing or a reusable voice or face that can make statements the creator never made | A separate likeness grant naming the product, the exact uses, the channels, the term, approval and a fee for the capability |
| Model training or fine-tuning | Use the creator's footage, voice, scripts or images to build or improve a model that can be reused after the deal | A separate training grant, or a refusal, plus retention and deletion terms that can actually be proved |
The copyright backdrop is narrower than many buyers assume. The U.S. Copyright Office concludes that outputs of generative AI can be protected only where a human author has determined sufficient expressive elements, and that this can include a human-authored work perceptible in the output or creative arrangements and modifications, but not the mere provision of prompts. Using AI to assist creation, or including AI-generated material in a larger human work, does not bar copyrightability. That matters to the licence because the brand may own or license a human-authored layer while the creator's likeness remains a separate right.
The regulatory backdrop is also thinner than the marketing suggests. Coblentz Law states that there is currently no general FTC rule requiring disclosure solely because AI was used. What the law does regulate is the impression the advertisement creates. A caption tool that leaves the creator's real experience intact is a different thing from a replica that puts a first-person product claim in their mouth. The grant should follow that difference.
Put ten terms in the AI clause
A usable AI clause reads like a short brief, not a sentence of legal fog. These are the ten fields a brand and a creator can actually negotiate, drawn from the checklists in Creators Agency, Snippet and Lathrop GPM. Leave one blank and the other side fills it with the broadest reading.
- Inputs: the final approved video, raw footage, outtakes, stills, audio, scripts, transcripts, name, voice and face, each named on its own line and not gathered under creator materials.
- Purpose: the specific business job, such as captioning one approved video, not marketing purposes in general.
- Outputs: what may be generated, from a translated caption to a synthetic scene, and what may never be generated.
- Training: a yes or no on training, fine-tuning, testing or improving any model, including a vendor's general service.
- Replica: a yes or no on a digital voice, face, movement or avatar, defined so derivative works alone cannot be read as consent.
- Approval: what the creator approves, how a request arrives, the response window, and what happens when the creator says no.
- Term and territory: channels, paid or organic use, markets, start and end dates, and any media-spend cap.
- Vendors: the named agencies, production partners and AI vendors, whether they may retain or sublicense, and who stays responsible.
- Deletion: source files, embeddings, trained models, prompts containing creator material, vendor copies and unused outputs, with written confirmation where the use is sensitive.
- Compensation: the AI right priced as its own line, separate from the content fee and from ordinary usage.
Two of those fields decide whether the deal is repairable. Deletion is harder than it looks: Creators Agency notes that proving training data, model weights, vendor copies, cached files and derivative outputs are gone is much harder than deleting a final video. Specificity is the other. Lathrop GPM warns that vague grants risk being unenforceable and also create avoidable friction when business teams later try to reuse content, which is a commercial loss on both sides.
Notice what the list does not do. It does not name a technology, because a clause that permits any tool now known or later devised prices a future nobody has scoped. It names the job, the file and the output, and those survive a tool change.
The ten fields that turn a broad AI clause into a scoped grant. Original diagram built from the Creators Agency, Snippet and Lathrop GPM checklists stored in the dossier.
Keep contract, approval, platform and advertising as four gates
An AI edit has to pass four separate checks, and the most common expensive mistake is treating one as proof of another. Creators Agency sets them out as separate questions: whether the contract grants the brand the right to create and use the output, whether the creator approved this exact meaning and context, whether the uploader applied the required platform label, and whether the endorsement and product claims are truthful and properly disclosed. The same page states the rule that matters: passing one gate does not pass the others.
The advertising gate is the one brands most often assume is handled by the label. It is not. The FTC Endorsement Guides require an endorsement to reflect the accurate experience and opinion of the endorser, and they say plainly that you cannot talk about your experience with a product if you have not tried it. Endorsements claiming specific results are usually read as meaning others can expect the same result. Those duties do not disappear because a tool wrote the sentence or because a hashtag sits beneath it.
Disclosure placement is a separate detail with its own rule. The FTC says the ultimate responsibility for clearly and conspicuously disclosing a material connection rests with the influencer and the brand, not the platform, and that a disclosure is more likely to be seen when it is very close to, or part of, the endorsement it relates to. A label in a bio, a comment or a platform toggle does not automatically meet that test. Keep the four questions on one page and answer each one before publishing.
State law adds a fifth pressure point without replacing the others. Snippet reports that under California and New York law a digital-replica clause in a services contract can be unenforceable under certain conditions, including when it has no reasonably specific description of the intended uses. It also notes that a vague description can still hold up if the uses fit the rest of the contract and the character of what was actually filmed or recorded. Coblentz Law reports that New York now requires advertisements that include a synthetic performer to disclose that fact conspicuously. None of this turns a label into a licence.
Four gates, not one. Original diagram built from the Creators Agency four-question frame and the FTC Endorsement Guides read for this page.
Price the AI right as its own line
There is no published market rate for an AI-editing or model-training grant, and any page that prints one is guessing. What the sources do give is a sense of scale for the neighbouring usage rights, which is why AI rights should sit on their own line and not hide inside a percentage. Netinfluencer reports that usage fees typically add 20% to 50% on top of a creator's base rate, and that perpetual rights can add 50% to 100%. The same piece records that usage premiums run about 20% to 30% for 30 days and up to two to three times the base for exclusivity.
Those figures are anchors, not an AI price list. A 30-day paid-usage uplift does not price a reusable voice model, because the model can generate work after the campaign and does not expire with it. Several practitioners in that piece make the split explicit: AI training and likeness modification belong in their own clause, priced on their own, and one calls AI training rights a license to a fundamentally different product, not a usage right. Treat the AI line the way you treat the content fee: a distinct deliverable with a distinct scope.
The negotiation also has a longer horizon than a campaign. LegalLens notes that copyright in the UK and US lasts for the creator's lifetime plus 70 years, so a perpetual AI grant runs far beyond any sensible campaign window. The same page says standard licence periods in the influencer industry range from 30 days to 12 months depending on scope. An AI right that is priced like a 30-day post but written like a perpetual assignment is mispriced before the first output is generated.
Price what the right can do, not the label attached to it. A caption tool on one approved video and a reusable likeness model are not points on the same rate card, and a buyer who bundles them will either overpay for the first or underpay for the second.
Repair a legacy grant before you rely on it
Many agreements in force today were drafted before AI tools were routine. The question is whether the old grant already reaches the new use. LegalLens describes the risk directly: if a contract grants a perpetual, irrevocable licence in any media now known or hereafter devised, and it was signed before AI clauses became standard, the brand may argue that AI training and digital twin creation fall inside that existing licence. The same page calls the point contested legal territory, not settled law.
The other side of the argument is narrower. Coblentz Law observes that a general grant of rights to use an influencer's name, image, likeness, voice and content may not clearly extend to a replica that generates new statements the influencer never made. Two careful readers, two different readings, which is exactly why a legacy clause should be reviewed, not assumed either way.
When the clause is already signed, keep three things apart. The licence may be enforceable, so the brand may have a legal basis to keep using the content. A licence, even a perpetual one, is not the same as a copyright assignment, so the creator may still own the work and may still license it elsewhere unless exclusivity prevents that. And a perpetual licence is only as broad as its other terms, so a use outside the named channels or territory is a breach question on its own. The Dealroom template makes the same distinction from the drafting side: a perpetual licence transfers value without transferring copyright, and ownership can stay with the creator, pass to the brand by assignment, or be shared through a licence.
For the next deal, replace the broad sentence with a specific, time-limited grant and write down the prohibited uses. For the current one, ask counsel which reading applies, because the sources give general information and not advice on a particular agreement.
Run the licensed output on one operated account
Once the grant is clear, the operating question is who puts the approved output into the world. A licensed AI edit that never ships has no value, and a publishing loop that runs outside the grant is a liability. This is the practical layer where a small brand usually needs help: someone has to keep the approved claims, the disclosure and the account permission in step while the daily work continues.
Superdirector is built for that layer. At $199 a month, one AI-run operated social account carries the daily loop of watching, drafting, preparing assets and publishing eligible posts under account-level standing posting permission, which is granted once per account after a warning that names the account, the public effect and the cost. That standing permission carries the account's eligible posts with no approval step on any single post. At $29 a month, the plan is MCP access and hands-on creation from inside your assistant: it drafts, plans and prepares, and you keep the button. The $29 plan does not post an account for you.
Neither plan decides the AI grant for you. The brand still writes the clause, the creator still approves the meaning and context, and a person still owns the disclosure call. What an operated account removes is the daily coordination cost of running the approved output, so the rights you negotiated are the rights that actually get used. If you are still choosing the tool, our page on how to read creator usage fees prices the neighbouring rights, and the contract skeleton shows where the AI clause sits among the rest.