Everything that breaks in AI-generated video ads

The failure mode people expect from AI video is obvious badness: melted hands, uncanny faces, physics that gives up halfway through. Those are real, and they are also the easy case, because you can see them.

The failures that actually cost you are the ones where the clip looks completely fine. Here are four we hit across 51 generated ads, in rough order of how long each took to notice.

1. The wrong ad, silently

One of our formats is a three-person podcast: three presenters in a row along the bottom of the frame, gameplay above. For several weeks it produced a single person on a sofa.

The job succeeded. The clip was well made. It simply was not the format that had been ordered, and nothing anywhere said so — no error, no warning, no failed status. It was found only because someone watched a clip and thought it looked different from the preset it was named after.

The cause was a preset that declared its panel count but not its style, and a code path that branched on style. It fell through to the single-presenter builder. The lesson is not about that bug specifically — it is that a generation pipeline has no natural way to notice it produced the wrong thing. Success means the vendor returned a file.

2. It says your product’s name wrong

An early clip for AllTrails said “AllTracks”. One for MyFitnessPal said “MyFoodiePal”, and another “MyFitnessPap”. The model approximates an unfamiliar compound word and lands one consonant away.

An ad that names the wrong product is unusable no matter how good it looks, and it is the single hardest defect to catch — everything else about the clip is fine, and you stop hearing the error after the second viewing.

When we went back and re-checked every clip we had ever generated, using honest transcription:

41% overall. Two-fifths of everything produced, naming the product wrongly or not at all.

There is a trap inside the fix, too. The obvious way to catch this is to transcribe the clip and compare against the app name. But if you also give the transcriber the app name as a hint — which every transcription API invites you to do, and which genuinely improves accuracy — it will hear the name you told it to expect. We did exactly that. The same audio transcribed as “MyFitnessPal” with the hint and “MyFitnessPap” without it. The detector had been blind since the day it was built, reporting clean on clips that plainly said the wrong thing.

3. It invents your interface

Ask for a shot of someone holding a phone and the model will draw something on the screen. It cannot render text — no image model reliably can — so what it draws is your brand name, misspelled, in an app that does not exist.

We watched a clip render a fitness app whose header read “MyFitnessPay”, with navigation tabs reading “Molca”, “Dirpiceas” and “Dialio”. In another, after we forbade text on the screen, it wrote the brand name on a chalkboard in the background instead.

That second attempt is the instructive one. Suppressing the behaviour in one place moves it somewhere else, because the model is not rendering text on a surface — it is putting the brand into the scene, and any surface will do. Six rounds of prompt wording did not fix it.

What does work is narrower than “tell it not to”:

• For games, let it draw the screen. A board or playfield has no wordmark to misspell.
• For everything else, frame past the problem — show the app scrolled into its content, where the title bar is simply off-screen. Charts and rows garble invisibly; a brand name does not.
• If the real interface matters, composite your actual screenshots in afterwards. A model will never render your UI correctly, and that is not a prompt problem.

4. It crops the presenter’s head off

Any format that puts a generated presenter over your footage has to cut that presenter out of their frame. Get the crop wrong and you ship three people visible from the eyebrows down.

Our first approach detected where to crop using black-bar detection, which is a proxy for where the people are. It fails in the specific case that matters: a clip where the presenters sit over a background that is not black reads as a full frame, falls back to an assumed position, and the crop lands across their foreheads.

The fix that worked was to stop inferring and start looking — a vision call on sampled frames reporting where the people actually are, with the crop following them. The same pattern as the format check in failure #1, and the same underlying lesson: the reliable way to catch a visual defect is to look at the pixels, not at a proxy for them.

One refinement matters: sample more than one frame. A clip that cuts between a wide shot and a close-up cannot be served by a single fixed crop, and one frame will never tell you that. We only learned this by shipping a clip that looked perfect at the midpoint and played its opening seconds as a headless torso.

There is a fifth failure worth naming here, though it is not the model’s fault: composing something important into the part of the frame the platform covers with its own interface. It has the same signature as the four above — the file is fine, the defect only exists in the feed. The exact percentages TikTok and Meta reserve are written up separately.

What the four have in common

None of these produced an error. Every one returned a finished, well-made video file. The pipeline had no way to know anything was wrong, because the only signal it had was “the vendor returned a file”.

Which is why the useful investment in generated video is not a better prompt or a newer model. It is a checking layer: something that looks at the output and asks whether it matches what was ordered, and refuses it when it does not. Every fix above that actually held was some version of that, and every fix that failed was an attempt to talk the model into behaving.

Method

Figures are from 51 ads generated for real published apps between August and September 2026, re-transcribed without name hints to measure naming accuracy honestly. Format percentages are small samples — some formats had only a handful of clips — and should be read as direction, not precision.