Case study
Why your AI-made health ad shows the wrong picture
An ad for a health supplement cut to footage of a different condition on about half its cutaways. What caused it, how we changed how the tool picks clips, and what an owner can check before an ad goes out.
Key takeaways
- A tool that chooses footage from written descriptions alone will use a poor clip rather than no clip.
- In a health ad, a single cutaway can make a claim the script never made.
- Blocking the wrong subjects is the first step; looking at the picture itself before the video is built is what catches the rest.
- A small library built for one kind of ad will fail an ad about something else.
- When no footage fits a line about the product, showing the product is the safer choice.
An ad for a health supplement cut away from the presenter to footage of a different health condition, on about half of its cutaways in the worst test ads. We changed how the client's AI video tool chooses each clip, and within three days of those first test ads the new clip picking was live. Replayed on the same ads, footage of another condition now appears on under one in ten cutaways.
What a wrong cutaway does to a health ad
In a health product ad the picture carries as much weight as the words, so a cutaway that shows something else makes a claim the script never made.
Footage of a different condition is the clearest case: the viewer reads it as a statement about what the product is for. Pictures of cash and luxury settings make a supplement look like a get-rich offer. Drinks, or a medical picture with blood in it, can get an ad turned down by the ad platforms before a customer ever sees it.
An editor can catch all of that, one clip at a time, by watching every cut of every ad. That is the work the tool was brought in to remove, which is why the pictures went unchecked.
Why an AI video tool picks the wrong footage
The tool was matching each clip on its written description alone, and when nothing was close it took the nearest clip anyway, however poor.
Three things made that go wrong at once:
- The team's own library held a few hundred clips, nearly all about one condition. An ad about a different condition had almost nothing to fit, so the nearest wrong clip won.
- The general library also held footage made for the client's other advertising, luxury settings among it.
- A description can share words with a line while the picture shows something else entirely, so a clip described in the right words could still show the wrong thing.
There was a separate fault on top. The shot list sometimes asked for a picture of the speaker, and the tool answered with a stranger's face talking to camera, sitting under the presenter's voice.
What we changed
We put three checks between the script and the finished video, so a clip can no longer reach the screen just because its description sat near the line.
The first reads the script, works out which health condition the ad is about, and rules out any clip showing another one. It also never takes footage of cash, luxury settings, drinks or medical pictures, nor a shot of someone talking to camera. A clip from the general library now has to match far more closely than before.
The second is a vision check: an AI model that looks at frames of the clip itself, rather than its description, alongside the line's words, the product and the audience. It answers four yes or no questions — another subject, another product, words on screen, the wrong person — and a failure counts only when a second look agrees. A failing clip gives way to the next one. When nothing in the library fits, the tool searches stock footage using the line's own words.
The third covers what is left. When a line is about the product itself and no footage fits it, the ad shows the product on screen instead.
All of this applies to the client's health supplement ads only; their other ads work as before.
What the business has now
The change is live in the client's video tool for every health supplement ad they make.
Replayed on the test ads' own cuts, footage of another condition fell from about half of the cutaways to under one in ten. Our engineer had already marked a set of cuts by eye; tested against those, the vision check failed every misfit and passed every good clip, on two runs. On the same cuts, about three in five library clips stayed where they were and the rest were swapped for a better one. Every clip of cash, luxury and the wrong subject was gone.
The client chose the condition filter for every health ad; the rest followed from what the test ads showed.
What you can check yourself
- Watch your next AI-made ad with the sound off and ask whether each picture fits the words on its own.
- Look at your own clip library and ask how many clips in it are about the topic of the ad you are making now.
- Check that no clip in a health ad shows cash, alcohol or a medical picture.
- Read the shot list before the video is built and see whether any line asks for a picture of the speaker, which is how a stranger's face ends up under your presenter's voice.
- Ask whoever makes your ads what the tool does when no clip fits a line.
What to do next
Take one ad you have already published and watch it once with the sound off, writing down any cut that does not fit its line. That list is the brief for whatever comes next: a tighter library, a rule that blocks a subject, or a check that looks at the picture before the video is built. Keeping the decision about what an automated tool may show, and proving it on your own ads, is the kind of work our engineers do.
Questions people ask
Why did my AI video tool put unrelated footage in my ad?
It matches clips by their written descriptions, and when nothing is close enough it takes the nearest one anyway. If your library is mostly about one topic, an ad about another topic has very little to fit. A description can also share words with your line while the picture shows something entirely different.
Can an ad platform reject my ad because of the pictures, not the script?
Yes. The platforms judge the whole ad, and the footage is part of what they review. In our experience, drinks and medical pictures with blood are what cause trouble in health advertising, even when the words are careful.
How do I check an AI-made ad before it goes live?
Watch it once with the sound off and look at each picture on its own against the words being spoken at that moment. Then read the shot list and ask what the tool does when nothing in the library fits a line. Both checks take a few minutes and need no technical help.
What should an AI video tool do when no clip fits a line?
It should move to the next candidate rather than settle for the nearest wrong one, and search stock footage using the line's own words. If the line is about the product itself, showing the product on screen is better than an unrelated picture.
Do I need a separate clip library for each kind of ad I make?
Not necessarily, but a library built mostly for one kind of ad will serve another kind badly. In this case, the team's clips were nearly all about one health condition, so an ad about another condition had almost nothing to fit. The practical step is to look at the picked clips against their lines before the video is made.