AI fits inside a production pipeline in four places: pre-visualization, script iteration, voice and translation, and supplemental b-roll. It does not fit in the hero footage, the on-camera performance, or the final cut. The rule we run every project by is simple: machines draft, humans decide, and anything the audience will judge the brand on is filmed, directed and edited by people. The title of this page names the real problem. The audience never sees the model. They see a human-looking result and judge it as if a person made it. If the output reads as synthetic, the viewer discounts the brand, not the software. Every decision below follows from that.
Where does AI go, and where does it stay out?
| Stage | AI role | Human role |
|---|---|---|
| Pre-visualization | Storyboard frames, shot-list variants, lighting and location mockups | Director picks, revises, and shoots the real thing |
| Script | First drafts, alternate hooks, cutdown versions | Writer owns voice, claims and final text |
| Voice and translation | Cloned VO for language cutdowns from one approved master | Talent consent, native-speaker QC, pronunciation fixes |
| B-roll | Inserts that cannot be filmed on budget: aerials, abstract textures, impossible scale | Editor decides whether it survives the grade |
| Hero footage, performance, final edit | None | Entire responsibility |
What does pre-visualization with AI change on set?
Before generative image tools, a storyboard for a two-day commercial shoot meant a day or two of an illustrator’s time and little room to iterate. Now a director can produce twenty frame variants in an afternoon, test three lighting directions for the same setup, and send the client a board that looks close to the finished grade.
The trade-off is expectation drift. A photoreal pre-viz frame gets mistaken for a promise. We label every board “reference, not final” and show at least one frame from the actual location scout next to the generated one. The upside is measurable: fewer setups abandoned on the day, and shorter shot lists because the weak ideas died in pre-production rather than on a paid crew’s clock.
How do we use AI for scripts without losing the brand voice?
Models are good at volume and bad at judgment. We use them to generate alternate opening lines, compress a 90-second script into 30- and 15-second versions, and flag sentences that run long for the on-screen timing. A writer then rewrites, verifies every factual claim, and cuts anything that sounds like a model wrote it.
The tell is uniformity: balanced clauses, hedged superlatives, and tidy summaries at the end of every paragraph. Humans speak unevenly. A script that reads too smoothly on the page usually plays flat on camera, and the talent will tell you so at the first take.
Is voice cloning and translation production-ready?
Yes, for cutdowns and localization, with two hard conditions. The first is written consent from the voice talent that covers cloning, each target language, the usage window, and compensation. The 2023 SAG-AFTRA television and theatrical agreement set the industry baseline for consent and payment on digital replicas, and non-union talent now expects the same terms. The second condition is native-speaker review of every language. The generation step costs very little. The review step is where the real budget goes, and skipping it is how a brand ends up with a mispronounced product name in six markets.
Typical practice: one English master with a human VO session, then cloned cutdowns in the additional languages. Lip-sync tools can re-time an on-camera speaker, but results on wide shots and fast speech are still inconsistent enough that we use them only on talking-head framing.
When is generated b-roll acceptable?
When the shot is not about the product, the people, or the place the brand is selling. Abstract textures, scale shots, weather, and establishing aerials that would otherwise need a permit and a drone operator are reasonable candidates. Current generative video models produce clips in the five-to-ten-second range, which is the right length for an insert and the wrong length for a scene.
What goes wrong: continuity between generated shots, hands and text rendering, logos that drift, and a plasticky surface quality that becomes obvious once the clip sits in a graded timeline next to real footage. Our working rule is that every generated clip must pass the colorist and the editor without either of them being told which clip it is. If someone spots it, it comes out.
Does the audience need to be told?
Increasingly, yes, and the obligations are now written down. YouTube requires creators to disclose realistic altered or synthetic content under its synthetic content disclosure policy. The EU AI Act, published as Regulation (EU) 2024/1689, sets transparency obligations for deepfakes and AI-generated media that apply from 2 August 2026, which covers any campaign running in European markets. In the United States, the FTC’s final rule on fake reviews and testimonials explicitly bans AI-generated testimonials, and its Endorsement Guides apply to synthetic spokespeople the same way they apply to human ones.
For provenance, we attach content credentials to delivered masters where the client’s platform supports them. The C2PA specification is the open standard behind those credentials and is already supported by Adobe, major camera makers and several platforms.
Who owns what the model produces?
Purely machine-generated frames are not copyrightable in the United States. The U.S. Copyright Office’s AI reports conclude that protection requires human authorship, and that prompts alone do not qualify. A generated b-roll clip dropped into a human-edited film is protected as part of the whole, but the clip on its own is not. That is one more reason the hero footage stays filmed: it is the asset the client actually owns.
What changed since this page was first written?
- Generative video gained native audio. Mid-2025 models began producing synchronized sound and dialogue with the picture, which made short inserts faster to review but did nothing for shot-to-shot continuity.
- Disclosure went from voluntary to required. Platform policies and the EU AI Act timeline turned labeling into a delivery item rather than a preference.
- Consent became contractual. Voice and likeness clauses are now standard in talent releases rather than an afterthought.
- Risk management got a framework. Larger clients now ask vendors to map their AI use to the NIST AI Risk Management Framework, so we document which stage used which tool on every job.
Does any of this affect search and distribution?
Google’s position, stated in its guidance on AI-generated content, is that it rewards helpful content regardless of how it was produced and penalizes content made primarily to manipulate rankings. The practical reading for video marketing is the same as for the footage itself: if the output is indistinguishable from careful human work and serves the viewer, the production method is irrelevant. If it is not, no amount of volume rescues it.
If you are planning a project that touches any of these stages, start by listing which assets the audience will judge the brand on and reserve those for live capture and a human edit. Then decide, per stage, where a model saves real time, what consent and disclosure it triggers, and who reviews the output before it ships. That document becomes the AI section of the production plan, and it is the one the client, the talent and the platform will all eventually ask to see.
