Where AI avatar faces come from: actors, photos and generated people

An AI avatar's face starts in one of three places

A visitor meets a friendly face on a call and wonders whether there's a person behind it. The honest answer depends on where the face came from.

Every vendor detail in this guide comes from that vendor's own pages, read on 6 October 2026. Across them, a face starts in one of three places:

• –Recorded: a real person filmed for it — an actor in a vendor's stock library, or someone at your company.
• –From a photo: a single still image of someone, animated into a face that speaks.
• –Generated: rendered from no one in particular, so it resembles nobody.

A stock face may be a real actor, and its terms limit what it says

Ready-made faces are the quickest way to start, and some vendors say outright that real people stand behind them. Synthesia, which makes avatar video, describes its Talent Experience Program as a community of the actors behind its most popular stock avatars. Beyond Presence says its stock avatars feature professional models filmed in studio settings. Anam's acceptable use policy refers directly to the actor behind a stock persona.

That reference comes with limits worth reading before anyone writes a script. Among its prohibitions, Anam's policy lists using a stock persona in a way that may attribute your statements, opinions or commentary to the actor, with discussions of medical conditions or disabilities as its example, and it asks for Anam's written consent before a stock persona appears in paid social advertising, political campaigns, fundraising or news.

Stock also means shared. Beyond Presence, for instance, describes its stock avatars as available to every account and a custom avatar as exclusive to yours. A face from a shared library may be greeting another company's visitors too, which matters if you meant the face to become part of your brand.

Whichever vendor you choose, ask what its terms rule out for stock faces, and read the answer against the words you plan to put in that face's mouth.

A custom face starts from someone's footage or photo, and vendors document their own steps

A custom face is trained on a real person: your head of sales, a founder, a presenter you hired. The vendors document how they handle that person's permission in their own ways.

Tavus trains a custom face from a photo or a video, and its documentation says the customer is responsible for having the rights and permissions for the likenesses it uploads. HeyGen's LiveAvatar asks for a consent recording for a video avatar, a short video documenting the subject's permission to create an avatar using their likeness. Beyond Presence says every avatar, from video or a single image, needs the explicit, informed consent of the person being replicated, and that it verifies this during review and creates no avatars of public figures. Anam's one-shot generator makes a speaking avatar from a single photo, and its acceptable use policy prohibits infringing a person's right to privacy or publicity.

What those steps have in common is the question they leave with you: whose face is it, and what did that person agree to?

A photo of someone stays a photo of someone

However a face is made, a likeness built from a real person's video or photo is still that person's. A generated version of a colleague is not an anonymous face; it's your colleague, speaking words they may never have said.

So agree the use in writing with the person whose face it is. Beyond Presence's own advice is a good outline: if the subject is a paid performer or an employee, make sure the agreement covers the creation and ongoing use of an AI likeness, the scope and duration of that use, and what happens if they leave.

Plan the exit before the entrance. An employee leaves, a contractor's engagement ends, a founder steps back: decide in the agreement whether the face retires with them, and who on your side switches it off.

None of this is legal advice. A question about a real person's likeness belongs with your counsel, and the answer can differ by country.

The voice and the source files are part of the likeness too

A custom face can bring a voice with it. Tavus's documentation says a face trained from video also creates a voice from the footage's audio, and Anam lets a team create a custom voice by cloning a recording. So the permission question covers how someone sounds as well as how they look, and the agreement should name both.

Then there's what the vendor keeps: the photo or footage a face was built from, and the versions generated from it. Anam's API, for one, documents a hard delete that permanently purges an avatar together with its source image and generated versions, intended for data-erasure requests, where an ordinary delete hides the avatar and keeps it restorable. Ask every vendor the same two things: where the source is kept, and how it's erased when the person asks or leaves.

And keep the marks on. Anam's acceptable use policy prohibits removing or disabling the watermarks and other mechanisms meant to tell AI-generated content from human-made content. A label is only worth something while nothing in the pipeline strips it.

A generated face resembles nobody, and it still carries its label

A fully generated face is rendered from no one in particular, so there's no person behind it to ask, to pay or to lose when they change jobs. It's still an AI speaking to your customers, so the label matters just as much as on any other face.

A generated face still deserves a look before it ships, the way you'd look at a stock photo. A face that happens to resemble a well-known person is a problem whatever produced it, and a realistic one can still be taken for a person on a call, which is why the label goes on it.

That's how Runner's own cast is made. Every character Runner ships is fully generated, with no actor pipeline behind it — no casting, no likeness licences, no footage of real people — and each carries the AI mark. Whatever name and title you give an agent, the words AI agent appear beside them unless your own wording already says it's an AI, and no setting takes them away.

A face your team uploads is different. A photo stays a photo of whoever is in it.

Settle these before a face speaks for you

Each one has an answer you can write down, and the written answer is the one worth keeping.

• –Is this face a stock actor, a custom face of someone we know, or generated from no one?
• –For a stock face, what do the vendor's terms rule out, and does our script stay inside them?
• –For a custom face, what did the vendor ask of the person on camera, and do we hold their agreement in writing?
• –Does that agreement cover ongoing use, its scope and duration, and what happens when they leave?
• –Does the face bring a cloned voice with it, and does the agreement cover the voice too?
• –Where are the source photo or footage kept, and how are they erased when the person asks or leaves?
• –Does the face carry an AI label in words wherever it appears?
• –Who on our side decides when a face is retired?