The silent-clip tell is gone
A finance director in Manchester opens a video message from the chief executive. The face moves, the voice is right, the lips match the words, and the instruction is to release a supplier payment before noon. None of it happened.
What changed this week. Meta put Muse Video into early preview, the first mainstream text-to-video model to generate clips with native, synchronized audio rather than silent footage a creator scores afterward. It ranks third on the public human-preference leaderboard for text-to-video. Days earlier, xAI marked Grok Imagine complete, adding image and short-video generation on its Aurora model across the Grok apps.
Why the audio matters. The tell most people used to spot an AI clip was sound: fakes were silent or dubbed out of sync. Native audio removes that tell. A generated spokesperson can now speak your script, in your executive's voice, with the lips in time. The model quality is not the story. The disappearance of the easy tell is.
Why this lands on your desk, not your feed
This is a provenance problem, not a creativity one. For an operator the useful question is no longer whether AI can make the video. It plainly can. The question is whether you can prove which videos are yours and real, and whether the people in your AI-made content agreed to be there.
Impersonation is the near-term cost. A convincing video of a named executive approving a wire, endorsing a product, or announcing a recall is now cheap to produce and hard to disprove after the fact. UK Finance and European fraud teams already log voice-clone approvals; synchronized video is the next escalation. Set a verification rule now: no payment or public statement acts on a video alone, without a second channel your staff already trust.
The bottom line. Provenance is shifting from a compliance footnote to an operating control. Content Credentials, the C2PA standard that cryptographically signs where a file came from, is the practical hedge: sign your own genuine video so a customer or journalist can verify it, and treat unsigned video of your brand as unverified by default.
The consent fight already started
The consent fight is here. Within hours of the Muse launch, users pushed back on Meta using their uploaded photos to train the model. That is the second liability: not the output, but the training and likeness rights behind it. In the EU, generative video sits under the AI Act Article 50 duty to label synthetic media, alongside GDPR consent for any real person's face or voice.
For a European business the practical exposure is your own use. If your marketing team generates a spokesperson, a customer testimonial, or a staff scene, you own the consent trail and the AI-content label, not the tool vendor. A generated face that resembles a real employee, or a voice trained on a real customer call, is where the fine lands.
What to do this quarter. Three moves, none of them technical. Write a one-page rule on when AI video may represent your brand and who signs off. Turn on Content Credentials for the video you publish. And add "did a real person consent, and is it labeled" to the same checklist that already asks "is it on brand." The capability is here early; the governance is the part you control.
Read next: Meta's Ad AI Launches Around Europe, Not In It | Meta Wants to Rent Out Its Spare AI Compute
