Consent, not deepfakes: how licensed AI likeness works

This article is general information, not legal advice. Laws and platform policies in this area differ by country and change often.

A deepfake of a person and a licensed AI version of that same person can be built with the same tools, trained on the same kind of footage, and look identical on screen. The technology is not what separates them.

What separates them is whether the person agreed, what exactly they agreed to, whether viewers are told, and what happens when something goes wrong. Those four things — consent, scope, labels and takedowns — are the whole difference. This article sets out what each should mean in practice, so you can tell a genuine licence from a nicely worded one.

A deepfake is defined by the missing permission

Regulators have landed on roughly the same idea. The EU treats realistic AI-generated or manipulated content depicting real people as something viewers must be told about, and the European Commission says “certain AI-generated content should be clearly and visibly labelled, namely deep fakes and text published with the purpose to inform the public on matters of public interest” (European Commission). Platforms draw the same line: TikTok’s advertising rules prohibit using “the likeness (audio or visual) of a public figure without their permission” in misleading contexts (TikTok Ads Policy).

So “licensed AI likeness” is not a softer word for deepfake. It is a claim that four specific things are in place. Each of them can be checked.

The clearest existing standard for consent comes from data protection law. Under the UK and EU GDPR, consent means “any freely given, specific, informed and unambiguous indication of the data subject’s wishes”, given “by a statement or by a clear affirmative action”. The UK regulator adds that people must have genuine choice, be able to refuse without penalty, and be able to withdraw easily at any time; that you must name yourself and any third parties relying on the consent; and that silence, pre-ticked boxes and inactivity do not count (ICO, What is valid consent?).

That is a good test for likeness too, whether or not data protection law applies to a particular deal:

  • Freely given. No bundling likeness rights into terms someone has to accept for something unrelated.
  • Specific. Consent to one brand and one campaign is not consent to a category.
  • Informed. The person knows what will be made, who receives it, and where it runs.
  • A clear affirmative act. A recorded, dated agreement — not an assumption from a past shoot.
  • Withdrawable. A real route to stop new uses.

Note what withdrawal can and cannot do. Pausing future licences is straightforward; pulling an ad that has already been bought and scheduled usually is not. That gap is an argument for short licence periods rather than for pretending revocation is instant.

2. Scope: what, where, when and how long

Consent with no boundaries is not consent to anything in particular. A licence should name:

  • the brand, the product and the campaign;
  • the script, or the approved claim set;
  • the placements and formats — social ads, websites, TV, out-of-home;
  • the countries where it may run;
  • the start and end dates;
  • whether voice is included as well as appearance;
  • whether it is exclusive, and what that is worth;
  • whether a personal testimonial is permitted, which should always be a separate yes.

Some lawmakers now require this specificity. California and New York have made contract terms that let a digital replica substitute for work a performer would have done in person unenforceable in certain circumstances unless the use is described with reasonable specificity and the person had legal or union representation (California AB 2602; New York S7676B). Even outside those states, “reasonably specific” is the right bar. Our creator guide goes through the same ground from the creator’s side.

3. Labels: tell the viewer, and let the file carry proof

Disclosure protects the audience, and it protects the person whose face is used — an unlabelled clip can travel a long way from the ad it was made for.

Three layers are available today:

Platform disclosure. YouTube asks uploaders to disclose realistic synthetic content that makes a real person appear to say or do something they didn’t do, and may also show a label on the player for photorealistic material (YouTube Help). TikTok asks creators to label realistic AI-generated content and requires an AI disclaimer on ads. Meta requires people to “disclose, using our AI-disclosure tool, whenever they post organic content with photorealistic video or realistic-sounding audio that was digitally created or altered” (Meta). We compare the three in detail in AI disclosure rules for ads on Meta, TikTok and YouTube.

Provenance in the file. C2PA describes Content Credentials as working “like a nutrition label for digital content, giving a peek at the content’s history available for anyone to access, at any time” (C2PA). Credentials travel with the file, so the label does not depend on anyone remembering to tick a box.

Law. The EU AI Act’s transparency rules add their own labelling duties, and under the Commission’s timeline those rules took effect in August 2026, with a voluntary Code of Practice on Marking and Labelling of AI-generated Content published to help. Check the current dates before planning around them.

A licensed likeness should carry all three where possible: a platform label, provenance metadata, and a licence reference that ties the video back to a specific agreement.

4. Takedowns: a route that works, before you need it

The honest test of any likeness system is what happens the day something appears that nobody approved.

Platforms now offer specific routes. YouTube says “if someone has used AI to alter or create content that looks or sounds like you, you can ask for it to be removed”, through its privacy complaint process; an individual must be “uniquely identifiable”, the complaint must come from them or their legal representative, and YouTube weighs factors including whether the content is realistic and “whether the content contains parody, satire or other public interest value” (YouTube Help, Protecting your identity). YouTube has also built likeness detection, which “helps creators find content on YouTube where their face appears to be altered or generated by AI” for verified, eligible creators over 18 (YouTube Help).

Law is arriving for the worst category first. In the United States, the TAKE IT DOWN Act became Public Law 119-12 in May 2025. It covers “digital forgery” — intimate depictions of an identifiable person created with software, machine learning or AI — and requires covered platforms, on a valid request, to remove the depiction as soon as possible and not later than 48 hours, and to make reasonable efforts to remove known identical copies (Congress.gov, S.146). Broader federal protection for likeness has been proposed rather than enacted: a digital replica bill was introduced in 2025 (S. 1367), a revised version followed in 2026 (S. 4591), and as of October 2026 neither has become law.

For a licensing service, the equivalent obligations are practical: a named contact, a stated response time, the ability to stop new licences immediately, a record of every licence ever issued, and deletion of the underlying face and voice data on request.

The capture data is part of the promise

A consent-first system also has to protect the raw material. Australia’s privacy regulator treats biometric information, “including facial images, when used for automated verification or identification purposes”, as sensitive information under the Privacy Act (OAIC). Photos, video and voice recordings used to build a likeness deserve that treatment: encrypted storage, no handing raw files to advertisers, no training beyond the person’s own likeness without separate permission, and real deletion.

A short trust test

Whichever side of the deal you are on, ask:

  • Can you see the recorded consent, and does it name this brand and this use?
  • Is the scope written down — script, placements, countries, dates, voice?
  • Does every output carry an AI label and a licence reference?
  • Is there a takedown route with a named contact and a response time?
  • Can the person stop new uses and get their capture data deleted?
  • Is a testimonial treated as a separate, explicit permission?

If the answer to any of these is “we’d have to check”, the thing being offered is not yet a licence.

Where Likeness sits

These four tests are what we are building Likeness around: verified adult consent recorded per use, scope set by the person rather than the advertiser, labelled output with a licence reference, and a takedown path that exists before anyone needs it. It is in development and not yet open. See trust and consent for the detail, or what this means if you are the person being licensed, or email [email protected].


General information only, not legal advice. Check the current version of any law or policy mentioned here, and take advice for your own situation.

Sources

Previous post
Next post