Case study
Why your AI video ads look fake, and the cut that fixes it
Two of a client's AI-made video ads looked machine-made to the editors who approved them, for two different reasons. We traced each one, and both now come from the same way of cutting an ad.
Key takeaways
- An ad that looks machine-made can come from a cutting rule as easily as from the tool.
- When every cutaway line shows people who exist in no clip library, the tool has no choice but to generate them.
- Cutting one generated clip into pieces, one per key word, makes the picture step back into frames the viewer has already seen.
- Splitting every shot at a fixed length can send the second half of a shot backwards, which is why a fixed length is a poor rule on its own.
- Starting from no generated cutaways and choosing a share when an ad needs one gives a team something concrete to check against.
Viewers have learned to spot generated footage, and an AI video ad that looks fake loses their trust within the first seconds. A client of ours in the United Arab Emirates makes video ads in volume: a presenter talks to camera, and the picture cuts away to other shots while the voice carries on. Two of their recent ads looked machine-made to the editors who approved them, and neither cause turned out to be the AI tool itself.
Why an AI video ad looks fake: the cutaways, not the voice
The cutaways are the shots the picture cuts to while the presenter's voice carries on, and they are where an ad reads as machine-made, more than the voice or the script. Two symptoms are worth knowing, because each has its own cause:
- every cutaway is generated, so nothing on screen is real;
- a shot jumps back to a moment already shown, as if a short clip had been played again from its start.
For a team making ads in volume, the first of those is the one that costs time. Every ad has to be watched and corrected by hand before it goes out.
The ad set to 'mostly real' that went out with 71 AI shots
One ad was set to 'mostly real clips' and went out with one real clip against 71 generated ones. The editors had asked for the opposite: a real clip every 7 to 15 seconds, and no generated filler at all. In a different ad, they saw shots that jumped.
Why 'mostly real' did not mean mostly real
The answer was in the script. That ad told a first-person story, so 20 of its 22 cutaway lines showed the story's people, and the platform always generates the story's people with AI, because those people exist in no clip library. The other two lines needed a clip that matched their exact words.
So 'mostly real clips' turned out to be a ceiling on generated shots rather than a share of real ones. A script with people in every line never came near that ceiling, and the tool filled each cutaway the only way it could.
Why the shots jumped
The jumping had a separate cause, and we found it frame by frame in the other ad. One generated clip covered a whole line, and each of the line's three to five key words took its moment from that same clip, back and forth. A run of frames played, then the picture came round again to a frame from earlier in that run, which is what the viewer sees as a jump.
A pacing rule made it worse. It split any shot longer than 2.5 seconds, so the second half of a 2.9-second piece was sent back to frames already shown.
The cut the editors wanted, and what the ads look like now
A new way of cutting is now the default for new ads, and it starts with the presenter. He carries the ad, and after every 7 to 15 seconds of him the picture cuts to a real clip from the client's own library that fits the ad's subject and mood, never the same clip twice.
- The cutaway goes to the presenter line that starts nearest that moment, and a presenter line stands between two cutaways.
- Each cutaway is one clip, or two on a line longer than 4.5 seconds.
- Where the team wants generated shots, the brief asks for their share as an exact figure between 0 and 100, with none as the default. The platform honours it as a count of lines, the story's people first and spread across the ad, and every other cutaway is a real clip.
- Where a generated clip is used, its cuts play forward as one run, never back, and a split lands only where the clip itself changes shot.
The first version of the new rhythm had faults of its own. It cut away every 18 seconds instead of every 7 to 15, because it waited out the time and then took the next line rather than the nearest one, and it pushed three to five clips into a single four-second line.
Re-cut with the corrected rhythm, the ad that had 71 generated shots came out with 9 real cutaways, 8.7 to 14 seconds apart, with the presenter between them. New ads start with no generated cutaways unless the team chooses a share, and ads already made lose their jumps the next time they are rendered. The new rhythm went live for the editors on 13 October, and the fix for the jumping clips followed the next day.
What you can check yourself
- Count the generated shots in your last ad against the real ones. If almost every cutaway is generated, the ad will look that way whatever the setting says.
- Play the ad at half speed where one shot runs under several words, and watch for a moment that comes round twice.
- Ask whoever runs your ad tool whether 'mostly real' is an exact share or only an upper limit.
- Look at your own clip library. If it holds nothing that fits your ad's subject and mood, the tool has nothing real to cut to.
- Ask your editor whether the cutaways illustrate each word or hold a rhythm.
If you take one habit from this, let it be watching your own ad at half speed before it goes out. Tracing an ad that looks machine-made back to one setting or one cutting rule is the kind of work our engineers do for clients.
Questions people ask
Why does my ad look AI-generated when it is set to mostly real clips?
A setting with that name can be a ceiling on generated shots rather than a promise of real ones, and the label does not always say which. It only matters when a script leaves room under it, that is, when some cutaway lines can be filled by real footage that already exists. When every cutaway line shows people the library does not hold, the ceiling is never reached.
Why does a shot in my ad repeat a moment I have already seen?
One clip has usually been cut into pieces, one per key word, with each piece taking its moment from the same clip and stepping back and forth inside it. A rule that splits every shot at a fixed length does the same, sending the second half of a long shot back to frames already shown. Playing the ad at half speed and watching the shots that cover several words is the easiest way to see it.
Should I use any AI shots in my ads at all?
That is a choice for your team, and the tools work either way. What is worth settling in advance is how the share is decided: as a figure somebody sets for each ad, not a style each editor interprets differently. A team that starts from none and asks for a share when it wants one has a clear default to check an ad against.
What clip library do I need so my ads do not look generated?
You need your own real footage, sorted so a cutaway can be filled with something that fits the subject and the mood of the ad at that moment. Without it there is nothing real to cut to, and the tool fills the gap with generated shots. It also helps if the library is wide enough that the same clip is not used twice in one ad.
Can ads I have already made be fixed?
The jumping clips go away the next time an ad is rendered, because that fix sits in the way clips are cut. The new rhythm applies to ads made from then on, so an older ad keeps the cutaways it was built with unless somebody re-cuts it.