An AI video pipeline for marketing: rights, consent and disclosure at every stage

Where the risk sits in an AI video

Every AI marketing video raises three separate questions, and they are answered at different stages. Do you have the rights to what goes in: faces, voices, footage, music, fonts and marks? What rights do you have in what comes out? And is what the video says and shows honest, including whether viewers can tell what was generated?

On the second question, the US Copyright Office concluded in January 2025 that, with current generally available technology, prompts alone do not give a user enough control to be the author of the output. Human contributions still count: your own footage and script, the selection and arrangement of material, and creative changes to generated output can be protected, judged case by case. In practice, the protectable parts of an AI video are mostly the human-made ones. A purely generated shot may be free for anyone to reuse, so do not build a brand mascot or a signature visual on it.

Terms used in this guide

• Likeness: a person’s face, body, voice or other recognisable traits.
• Digital replica: the Copyright Office’s term for a realistic, digitally generated version of a real person’s voice or appearance.
• Synthetic content: footage or audio generated or materially altered by AI so that it looks or sounds real.
• Licence scope: what a licence allows, by channel, paid or organic use, territory and duration.
• Asset ledger: one row per asset in the video, recording its source, licence or consent, scope and whether it was generated.
• Content Credentials: provenance data attached to a file under the C2PA open standard, recording its origin and edits.

The pipeline, stage by stage

Stage 1: brief — decide what may be said, and who may appear

The brief sets the limits every later stage checks against. It lists the approved claims, each traced to a source you can show. It names every real person who will appear or be heard, so consent can be collected before anyone writes a script around them. It lists the channels and markets, because those decide which licences you need and which disclosure rules apply. And it records the disclosure stance: platform labels where required, plus whether the video carries its own on-screen line.

Stage 2: script — keep claims and voices honest

The FTC’s rule on consumer reviews and testimonials, in force since October 2024, prohibits fake testimonials, including ones attributed to people who do not exist. The FTC’s own questions and answers on the rule add two points that matter for AI video. The rule has no blanket ban on AI avatars, but an avatar delivering a testimonial is only acceptable when the testimonial itself is real and not misrepresented. And using a celebrity’s avatar without permission violates the rule when viewers would think the celebrity gave the endorsement. At script stage, that means every claim traces to the approved list, no line invents a customer, and no presenter speaks as if they personally used the product unless they did.

Stage 3: storyboard — run the rights check before render

Rendering is where cost accumulates, so clear rights while frames are still drawings. Give every asset in every frame a row in the ledger, and block the render while any row is incomplete. The ledger is also what you hand a platform, a client or a lawyer who asks where something came from.

frame,asset,source,licence_or_consent,channels,paid_use,territory,expires,generated,disclosure
01,founder presenter (face + voice),own talent,release R-2026-014,web+social,yes,worldwide,2027-10-31,yes,platform label + end card
02,inbox footage,stock library,licence SL-88213,web+social,yes,worldwide,perpetual,no,none
03,product screen recording,own product,n/a (owned; customer data blurred),all,yes,worldwide,n/a,no,none
04,customer review quote,customer email,written permission P-0042,web+social,yes,worldwide,2027-04-30,no,none
05,city skyline b-roll,image-to-video tool,tool terms v2026-09 (commercial use),web+social,yes,worldwide,n/a,yes,platform label
06,music bed,music library,sync licence ML-5530,web+social,yes,worldwide,2027-10-31,no,none

Stage 4: render — keep provenance intact

Before each project, re-read the terms of every generation tool you use: whether commercial use is allowed on your plan, who owns the output, and whether generating real people is prohibited. Keep the project files, prompts, tool versions and dates, so you can show how each shot was made. Where your tools support Content Credentials, keep them attached: C2PA is an open standard for recording a file’s origin and edits, and YouTube and TikTok both say they may label uploads automatically when that metadata is present. Re-encoding or cropping in some editors strips it, so check the exported file.

Stage 5: review and publish — disclose where the rules ask

Disclosure rules differ by platform and by market, and they change often. The table summarises what each source said when we read it; check the current text before each campaign.

Likeness and voice consent: what the release should cover

In the United States, protection against unauthorised digital replicas comes from a patchwork of state privacy and publicity laws, and the Copyright Office’s July 2024 report found those laws insufficient and recommended a new federal right. Until the law settles, a written release is the control you can rely on. A release for an AI presenter or a cloned voice should state:

• Who is being replicated, and exactly what: face, body, voice, name, or a combination.
• That the release covers AI-generated performances, not only recorded footage.
• Which scripts or categories of content are allowed, and which are excluded, such as endorsements, political content or anything the person has not seen.
• Channels, paid or organic use, territories and an end date.
• Whether the person approves each script, or a category in advance.
• Who holds the trained voice or avatar files, where they are stored, and when they are deleted.
• How the person can withdraw consent, and what happens to videos already published.
• Payment, if any, and what happens when an employee leaves the company.

Employees deserve particular care. Consent given as part of a job can look less than free, so keep the release separate from the employment contract and make withdrawal easy.

Worked example: what the rights check caught

The storyboard at the top of this page, as a rights check might go. It is an illustration with an invented product, not a client project.

1. Frame 01 passes: the founder’s release names this script, social and web channels, and an end date.
2. Frame 02 fails: the stock licence covered web use only. A clip with a paid-social licence replaces it before render.
3. Frame 03 needs work: a customer’s email address is visible in the screen recording, so it is blurred.
4. Frame 04 changes: the customer withdraws permission for ads, so the frame shows a product feature instead of the quote.
5. Frame 05 passes with a flag: the generated skyline is marked for the platform label.
6. At publish, the AI labels are set on each platform, the end card carries the agreed wording, and the ledger is archived with the project.

Rights, consent and disclosure checklist

• Every claim in the script traces to an approved source.
• No testimonial, review or endorsement is generated, and no presenter speaks as a customer.
• Every real person who appears or is heard has a signed release covering AI generation, scope and an end date.
• No public figure, competitor or private individual is replicated without written permission.
• Every asset has a ledger row with its licence or consent, channels, paid use, territory and expiry.
• Generation tools’ terms were re-read for this project and allow commercial use of the output.
• Project files, prompts, tool versions and dates are kept with the video.
• Content Credentials survive export where your tools support them.
• Each platform’s AI label is set wherever its rules ask, and an on-screen line is added where the brief requires one.
• Rules for each target market, including the EU and India where relevant, were checked on the day of publishing.