What happened
The Verge reported on July 17 that TikTok is starting to test an opt-in tool that scans for AI likenesses and lets creators report possible unauthorized use to the company. The report says the test is initially available to some U.S. creators.
According to the same report, participating creators first verify their identity through Jumio with a real-time selfie scan and an ID check. TikTok spokesperson Zachary Kizer told The Verge that TikTok does not retain ID documents, and that facial information is used only for likeness matching and to help identify potential unauthorized uses of a creator's likeness.
After verification, TikTok's system scans for AI-generated content that may use the creator's likeness. The creator can then review potential matches and report unauthorized posts or accounts. That is the reported product test. It is not a full public rollout, and it is not a guarantee that every misuse will be found.
The policy context
TikTok already tells creators to label content that is completely generated or significantly edited by AI, and it requires labeling for AI-generated content that contains realistic images, audio, or video. Its Help Center also says TikTok may automatically apply AI-generated labels when it identifies qualifying content or reads attached C2PA Content Credentials.
TikTok's Help Center draws a separate line around harm and consent. It says some AI-generated content is not allowed even if it is labeled, including the likeness of young people under 18 and the likeness of adult private figures used without permission. It also tells users to report AI impersonation through its deepfakes, synthetic media, and manipulated media reporting path.
The likeness test sits inside that wider platform move: labels, provenance signals, detection models, and user reporting are being stitched together. TikTok's March newsroom update also described invisible watermarking and C2PA-based labeling as part of its AI-generated content transparency work.
Why this matters for creators and experts
A likeness tool is about more than celebrity deepfakes. For any person whose credibility depends on their face, voice, work, or point of view, synthetic impersonation creates two problems at once.
The first problem is misuse. A generated video can make a person appear to endorse a product, explain a position, repeat a claim, or participate in a trend they never touched. Even a good detection tool is downstream of publication. The false version may already have shaped what viewers think.
The second problem is ambiguity. When audiences know synthetic media is common, they may become less certain about real content too. A founder's real product explanation, a consultant's real framework, or a researcher's real caveat can be pulled into the same trust fog as AI-generated imitations.
That is the REC angle. Detection is useful, but creators also need evidence of authorship before there is a dispute. They need a clear record of what the person actually said, what the team approved, and what later assets were derived from that source.
REC's read: keep the person upstream
AI likeness detection treats the person as something to protect after publication. A research-guided interview treats the person as the source before publication.
That order matters. If a team starts with a generated post, generated script, generated avatar, or generic prompt, it may produce something fluent without proving that a real person supplied the judgment. If the team starts with a recorded answer, the later article, clip, caption, or design has a source to point back to.
The value is not only defensive. A source-of-record video helps an editor preserve nuance. It shows the example the person chose, the caveat they added, the phrase they would actually use, and the point where they refused to overstate the claim. Those details are hard to reconstruct from a polished AI draft.
The same principle applies when AI helps prepare the conversation. AI can research the context, organize questions, transcribe the answer, and surface candidate highlights. The authorship still lives in the recorded answer and the human review that follows.
A practical workflow
If you publish expert video, product commentary, founder-led content, or creator education, treat identity and authorship as workflow requirements, not just legal cleanup.
First, record the claim before packaging it. Ask the speaker one focused question and capture the answer on camera. For example: what did you build, what changed your mind, what evidence supports this point, what would be misleading to imply, or what should the audience do differently?
Second, preserve the transcript and the approval trail. A transcript lets the team trace the final sentence back to the original answer. An approval note records that the named person reviewed the implication, not just the wording.
Third, label AI assistance proportionately. If AI helped prepare questions, organize a transcript, draft a caption, or edit the clip, say so where audience expectations or platform rules call for it. Do not let the label do all the trust work. The stronger trust signal is still the real person's sourced answer.
Fourth, separate real footage from synthetic presentation. If an AI-generated asset uses a real person's face, voice, or endorsement-like context, consent and clarity should be explicit. If the asset is only summarizing or designing around a recorded answer, keep the link to that source visible to the internal team and, when useful, to the audience.
The takeaway
TikTok's test is a useful sign of where platform safety is going. YouTube already has its own likeness detection help flow, and TikTok is now testing a comparable idea for some creators. The direction is clear enough: platforms are trying to give people more ways to find and report synthetic uses of their identity.
That work matters, but it does not replace creator discipline. Detection can help after something appears. A source-of-record workflow helps before publication, during editing, and after a question arises.
For REC, the rule is simple: let AI help with preparation and organization, but keep the person upstream of the claim. In an AI feed, the strongest content is not merely content that avoids impersonation. It is content that can show where the human judgment came from.