TikTok bans AI likeness of private people
TikTok bans a realistic AI likeness of a private adult without permission, and of anyone under 18 even with it. A label does not make the video allowed.
Nearly every AI content policy anyone has read in the last two years works the same way: make what you want, disclose it, and the disclosure resolves the problem. Two years of platform announcements have trained the reflex that labelling is the answer. TikTok has a category where it is not the answer, and where reaching for it means you have already lost.
TikTok's Help Center page on AI-generated content states that some AI-generated content is prohibited even if properly labelled. The prohibited set includes the likeness of anyone under 18, and the likeness of adult private figures used without their permission. A label is not a substitute for permission, and permission is not available as a cure for anyone under 18. This is a rule about whether the content may exist on the platform. The disclosure machinery sits entirely downstream of it.
A gate, not a label
The distinction is structural and it changes where the decision belongs in your process.
A labelling rule is a publishing decision. You make the asset, then you decide what to attach to it. If you get it wrong you fix the attachment. The asset is fine.
A prohibition is a generation decision. If the output is a realistic synthetic depiction of a real private individual that the platform does not allow — an adult used without permission, or anyone under 18 — there is no attachment that makes it publishable, and the money and time that went into producing it are spent. The only place to catch it is before you generate, which is upstream of every part of the compliance process most teams have built.
That is the whole practical point. Everything else in this post follows from moving the check earlier.
TikTok is not doing something eccentric here. It also reads C2PA Content Credentials on upload and applies its own invisible watermarking, and it sits on the C2PA steering committee. The prohibition exists alongside all of that. Provenance answers "what is this." The prohibition answers "may this exist here."
The line between private and public, and why you should not stand on it
The rule turns on the person being a private individual. Public figures are handled under a different set of rules — restrictions around endorsements and political content rather than a flat prohibition.
Do not build your workflow on the assumption that you can classify people accurately. The categories are the platform's, and the boundary is hard: a person with 40,000 followers, a local business owner, an employee in company marketing, a semi-known figure in one niche. Getting it wrong is a removal, not a labelling error.
Treat everyone as private unless they are unambiguously a public figure, and note that "unambiguously public" only moves you into a different set of restrictions. For a private adult, TikTok's published rule is permission, not a label. For anyone under 18, it is a ban. Neither path leads to "generate and disclose later."
A great deal of likeness thinking is calibrated on fame, because right-of-publicity law historically attached to the commercial value in a well-known identity. Regulation has been moving the opposite way. Denmark's copyright amendment, as described in the European Parliamentary Research Service briefing, is drafted to cover all natural persons rather than only recognisable ones. The instinct that says "nobody knows who this is, so there is no claim" is aging badly.
Consent-first, in the order the steps actually have to happen
The failure mode is a pipeline where consent is a publishing gate. Someone checks for a release just before the asset ships. That ordering fails against a generation-side prohibition, because by then the prohibited thing has been made.
Reordered:
- Classify the input before you generate. For every asset, one field: does a real, identifiable person's likeness or voice appear anywhere in the inputs? Reference photo, face scan, voice sample, a prompt that names a person. Answerable at brief stage; unanswerable six weeks later.
- Get consent covering the specific use, before generation. Not a generic model release. What is being generated, in what contexts it may appear, for how long, and whether it may be modified. Usage rights in creator contracts is the practical drafting version.
- Bind the consent to the asset, not to the project. One consent record, one list of asset IDs it covers. A consent that covers "the Q3 campaign" answers nothing when a specific clip is challenged.
- Default speculative work to invented presenters. Concept tests and pitch mock-ups should use a presenter who is not a replica of a specific real person. Knowing whether a given asset is a stock avatar, a digital twin or a generated character is the entire audit.
- Keep a revocation path. People change their minds, leave companies and die. If you cannot answer "which live assets contain this person" in one query, you cannot honour a withdrawal.
The table version, for pasting into a brief:
| What is in the input | What you need before generating | Publishable on TikTok |
|---|---|---|
| No real person | Nothing extra | Yes, subject to normal disclosure |
| A generated character resembling nobody specific | Nothing extra | Yes, subject to normal disclosure |
| A real person, filmed, not replicated | Normal talent release | Yes |
| A real private adult, realistically replicated, with documented permission | Permission matching the specific use | Conditional — still subject to how TikTok classifies the person |
| A real private adult, realistically replicated, without permission | Nothing cures it | No |
| Anyone under 18, realistically replicated | Permission does not help | No |
| A public figure, realistically replicated | Consent plus category-specific restrictions | Restricted, not open |
Note the under-18 row. A signed release answers a legal question and does not answer TikTok's rule for young people. For a private adult, the Help Center's own words are "used without their permission," so documented permission is the platform requirement, and a label is not a stand-in for it. Classification as private versus public is still the platform's call.
The voice half, which almost nobody checks
Likeness tooling is overwhelmingly built around faces. Detection systems look for faces. Consent forms describe appearance. Review processes screen video frames.
Voice is the gap that is not covered by any of that, and it is legally the faster-moving half. Tennessee's ELVIS Act added voice as a protected personal right effective 1 July 2024, including exposure for tool providers. California's AB 2602, effective 1 January 2025, makes a replica clause in a performer contract unenforceable for new performances where the intended uses are not reasonably specifically described and the performer was not represented by legal counsel or a union. At federal level the TAKE IT DOWN Act's platform duties took effect on 19 May 2026: covered platforms must remove nonconsensual intimate imagery, including digital forgeries of it, within 48 hours of a valid request, and FTC enforcement began the same day. That is a narrow intimate-image regime, not a general federal likeness right.
What is not law, and is routinely written about as though it were: the NO FAKES Act. It advanced out of Senate Judiciary on 18 June 2026 as S.4591 and remains a bill. There is still no federal right of publicity in the US, which is why the answer to "is this allowed" is genuinely different across state lines. Voice likeness law for creator marketing covers that terrain.
If you clone voices at all, run the same five-step ordering above on voice inputs. A cloned voice of a private individual is the same object as a cloned face for the purposes of the rule at the top of this post.
FAQ
Does adding an AI label make a synthetic depiction of a real person acceptable?
Not by itself. TikTok's Help Center treats some AI-generated likeness as prohibited even when labelled: anyone under 18, and adult private figures used without permission. Disclosure operates on content that is already permitted.
What if the person consented in writing?
For an adult private figure, written permission is what TikTok's Help Center requires, and a label is not a substitute for it. Consent still has to match the actual use, and it does not help if the person is under 18, if TikTok classifies them as a public figure in a restricted context, or if a separate legal right (publicity, contract, a state replica statute) is still open. Plan for the possibility that a fully consented, fully lawful asset is still not distributable on a given platform, and decide that before production rather than after.
Does this apply to a licensed avatar of a real person?
A licensed likeness is a consented replica, so the legal side is addressed. The platform side still depends on how TikTok classifies the depicted person — private adult, public figure, or under 18 — and on the licence terms themselves. TikTok's own avatar likeness licensing is worth reading because it shows the consented path for avatars the platform actually offers.
How do I find out if someone has made a synthetic version of me or a client?
Likeness-detection enrolment is the practical tool, and it works best for people whose likeness you legitimately control and want to monitor. Enrolling in likeness detection walks the process. It is face-oriented, which brings you back to the voice gap above: for now, monitoring for a cloned voice is meaningfully harder than monitoring for a cloned face.