We are AI-native and we still pay actors. Here is where each one wins.
After a year of running generative tools alongside a roster of real creators, an honest account of what each is genuinely better at, and the line we will not cross.
This piece is now part of a longer guide: UGC ads for performance: the complete guide.
We run generative video in production, and we keep a roster of more than eighteen hundred rights-cleared human creators. Those two facts are usually presented as a contradiction. They are not. They are a division of labour, and after a year of doing both we have a fairly clear view of where the line sits.
Where the generative tools genuinely win
One person talking to camera. This is the surprise. Single-speaker monologue with lipsync is the thing the current models handle best, and the emotional range you can get out of a well-written prompt is far beyond what the category was doing eighteen months ago.
Camera movement. Pans and tilts come out clean and cinematic with very little coaxing. Getting the same move out of a phone in someone's kitchen takes a tripod, a retake and a patient creator.
Variation at the edges. Once a concept works, generating the fifteenth framing of it is close to free. That is the part of production that used to eat the schedule.
Where humans are still not close to replaceable
Anything that has to be believed as testimony. The moment a piece of creative implies a real person had a real experience, a synthetic performance is the wrong tool, and not only for legal reasons. Audiences are getting better at spotting it every month.
Hands, products and physical interaction. Anything where the product is handled, opened, applied or worn. The models are improving and they are not there.
Unscripted specificity. The detail a creator adds because they actually use the thing. You cannot prompt for the detail you did not know to ask about.
The line we will not cross
No synthetic replicas of real people. No fabricated testimony, no invented endorsement, no fictional patient story presented as real. If a piece of creative implies a person exists, that person exists, they were paid, and they signed.
This is not a marketing position we adopted because it sounded good. It is a production constraint that shapes what we bid on, and some work does not fit inside it.
It also has a commercial reality attached: large rights holders increasingly restrict generative work contractually, so for a meaningful share of the accounts we run, the question is settled before we get to have an opinion about it.
What this means in practice
Generative tools sit in the middle of the pipeline, not at the front and not at the end. The idea still comes from a brief written by someone who read the performance data. The believability still comes from a human being who was actually there.
The teams getting this wrong in both directions are the ones treating it as an identity rather than a tool. Refusing the tools entirely means producing a tenth of the volume the auction needs. Using them for everything means producing a great deal of content that nobody believes.
Volume is not the hard part any more. Being believed is.