How do I prompt AI video tools so ad footage doesn't look like AI?
Short answer
Ask for handheld, slightly jerky camera work instead of slow, smooth motion, and direct cuts instead of transitions. Vary the people's ages, hair, and ethnicity so they don't all look like siblings. Then treat the output as a library of 3-second clips you mix and match, not finished ads.
The tell for AI video is the camera. Everything is very slow-motion-y, steady, centered. That's kind of a hallmark of AI.
What camera direction should I give?
"Think about it more as like a handheld camera, right? Someone there, kind of with slight jerkiness, kind of moving around to get the shot."
Not a steadicam gliding around the subject. Almost like the movie The Bourne Identity. The movement makes it harder for a viewer to fixate on whether a face is real.
Should the footage match the hook?
Yes. My event ads are just video of something happening around town with a line like "do you not want to miss events?" When someone showed me an AI video tool, my advice was to prompt for footage that makes the same point: a fun outdoor concert, a camera moving through a crowd, a festival in a park, outside not inside, a little faster than normal speed but not that fast.
Now when your ad says don't miss events, the video says it too. A recognizable landmark is the wrong pick. It's recognizable for those people, which is good, but it's not related.
What about transitions?
Ban them. One generator added a transition gimmick between every clip. Fancy transitions read as AI. Tell it to cut directly from one clip to the next.
Why do all the AI people look the same?
Because the model defaults to a type. When I generated footage for a senior audience, every person had the same white hair and looked like brother and sister. Fix it in the prompt:
- Vary hair color. A few people can have white hair, not everyone.
- Vary ages within the audience.
- Include a realistic mix of ethnicities for the target demographic.
- Ask for whatever's normal for that audience, not a uniform cast.
How do I keep AI people from being noticed?
Keep them on screen briefly. If you only see the person for one or two seconds, then cut to B-roll and back and forth, it's way harder to notice. Camera motion helps the same way: the few people who might try to study a face can't, because things are moving around.
How should I use what the tool generates?
As raw material.
"Think of these as a bunch of three-second clips that you can mix and match in any way you want."
- Run several prompts.
- Keep the good clips. Throw away the bad ones (we tossed a surfing scene that didn't fit).
- Aim for around 20 usable three-second clips.
- Mix and match them into many ad variants.
How polished does it need to be?
Less than you think. Of one batch of senior lifestyle clips, I said it was good enough to literally run, even if it's not perfect. Our plain, unbranded event ads have worked well enough that I don't overthink it. That annoys me, because I want to do very professional branding, and it just doesn't work.
A prompt skeleton I'd start from:
About 3 seconds per shot.
Handheld camera with slight natural shake.
No slow motion, no smooth gliding moves, no transitions between shots.
Scene: [local event that matches the ad's hook], outdoors, slightly faster than normal speed.
Crowd of [audience] with varied ages, hair colors and ethnicities.
People in frame in the first two seconds.
Whether you should use AI footage at all is covered in using AI video in Meta ads.
Answered by TJ Larkin. From creative working sessions with my ad team and AI tool demos, 2026. Last updated .