Rights, Likeness and Consent in AI Video: What You May and May Not Make

This is the part of the category that people skip and then discover. An AI video operation is unusually easy to start and unusually easy to lose: the production risk is near zero, and essentially all the real risk sits in what you generated of whom, and what you had the right to do with it.

The good news is that the rules are short and mostly intuitive. The bad news is that they are enforced after you have built something, not before.

Data Notice: This is a general orientation, not legal advice. Rights in likeness, copyright and AI-generated media vary substantially by country and are changing quickly. The platform’s own terms are the authoritative statement of what is permitted there, and a lawyer in your jurisdiction is the authority on the rest. Consult qualified professionals before making commercial commitments.

The two rules that cover most situations

Both are the vendor’s own stated position, and both are worth memorising:

Likeness requires authorization. Creating with a real person’s likeness is permitted only where that likeness is owned, licensed or otherwise approved for use. These platforms are built for authorized creators, brands and talent — not for unauthorized lookalike output.

Commercial use depends on your rights to the inputs. Whether you may sell or monetize the result depends on your rights to the source media, the avatar, any brand or product in shot, and the final content. If any link in that chain is unclear, the whole chain is unclear.

Notice what neither rule says. Neither says “if the tool lets you do it, it is allowed.” Technical possibility and permission are unrelated, and the platform’s content checks are a floor rather than a legal opinion.

Likeness, specifically

The situations, from clearly fine to clearly not:

Situation	Position
Your own face, your own photographs	Fine
A person who has agreed in writing, for this use	Fine
Talent you have a signed licence with	Fine, within the licence
A colleague or friend who said “sure” verbally	Weak. Get it in writing
A public figure or celebrity	Not permitted without authorization
Someone from photographs you found online	Not permitted
A composite “inspired by” a real person	Risky, and judged on resemblance

The last row catches people out. A character built to look substantially like a specific real person is a likeness question regardless of how it was assembled, and “it is not literally them” is not the defence it feels like.

Training is a bigger ask than a photograph. Training an avatar on someone’s photographs does not use those images once. It builds something that can generate new pictures of that person indefinitely, in situations they never saw and cannot anticipate. That is substantially more than consenting to one image being used, and treating them as equivalent is how people end up in disputes they cannot win.

Platforms record your confirmation of consent when you start training. That record protects the platform. It does not, by itself, protect you — what protects you is actually holding the consent.

Not a contract; a short document that removes ambiguity later. Six points:

Who is being depicted, and who is granting permission.
What is being created — an avatar trained on their photographs, and images and video generated from it.
Where it may appear — named platforms, or “public social media”.
Whether commercial use is included. This is the one most informal agreements omit and most disputes turn on.
How long it lasts, and how it can be withdrawn.
Date and signature. An email reply saying yes to a message that contains the above is generally better than nothing.

If the person is being paid, this belongs in the payment agreement. If they are not, it matters more rather than less.

Source material you did not make

Everything you feed in carries its own rights.

Your uploads. Your own pictures and video, and material you have the right to use. Uploads pass the same content checks as anything else.

Saving video from a link. Tools that pull a video in from TikTok, Instagram, YouTube, X or Facebook do not check whether you have the right to use what you point them at. That judgement is yours, and getting it wrong is a real problem rather than a technical one.

Reasonable: your own videos from your own accounts; material you have permission to use; material whose licence permits your intended use.

Not: someone else’s work republished as your own; anything the platform’s terms forbid saving; anything you intend to publish without the right to do so.

Republishing someone else’s video as your own is not a grey area. It is also the kind of thing that gets noticed, since the original creator’s name travels with the file.

Music. The most common accidental infringement in this whole pipeline. Platform copyright systems are automated, unsentimental and retroactive — they will find it, and they will take the revenue. Use licensed libraries or original music.

Brands and products in shot. A recognisable logo turns your video into something a brand’s lawyers may have views about, particularly if you are selling it.

Disclosure: label synthetic media

Most major platforms now require AI-generated or substantially synthetic media to be labelled, and several down-rank content that should have been labelled and was not.

Label it. The honest reasons are also the practical ones:

It costs nothing and there is no audience penalty worth the risk.
Failing to label is an account-level problem, not a video-level one.
Requirements are tightening rather than loosening, and content published today is still there when they do.

Where AI depiction of real people is involved, disclosure obligations are usually stricter and sometimes statutory.

Where the platform’s own moderation sits

Published work is content-rated automatically. Images are rated at once; video is reviewed after publishing because the check samples the actual clip rather than a single frame — a video whose first frame is harmless is not necessarily harmless. If a rating is uncertain, a post stays pending rather than going public: unreviewed is not treated as approved.

Posts can also be removed following a report or a rights claim.

Understand what this system is and is not. It is a filter on what the platform will host. It is not a determination that you had the rights, and passing it is not a defence if you did not.

Client work raises the stakes

Everything above gets sharper the moment someone is paying, because now there is a second party with an expectation and possibly a contract.

Settle before you start:

Whose rights cover the inputs — if the client supplies product shots or talent photographs, get their written confirmation that they hold the rights and are granting them to you for this work.
Who owns the output, and whether you may use it in your own portfolio.
Whether AI generation is disclosed to the client. It should be, in writing. A client who discovers it later, from someone else, is a client you have lost regardless of the quality.
What happens to the avatar if the relationship ends.

None of this needs a lawyer for a small engagement. It needs an email that says these four things and a reply that agrees.

A short pre-publish check

Before anything goes out, especially anything commercial:

Every person depicted is you, licensed, or has written consent.
No public figure or celebrity likeness without authorization.
Every uploaded input is yours or permitted.
Music is licensed or original.
No recognisable third-party brand you have not cleared.
Synthetic media labelled per each platform’s rules.
For client work: rights and ownership agreed in writing.

Six of the seven take seconds. The one that takes real effort — written consent — is the one that protects the operation you are building.