Guide
What makes AI video look cheap, and how do you avoid it?
AI video looks cheap mostly because of what surrounds the footage: a script with no point, text drawn inside the generated scene, faces and props that change between shots, the same framing in every shot, uneven sound and template edits. Avoid it by starting from one specific idea, putting every word and number in graphics layered on top, building recurring characters from reference photos, changing the framing at every cut, pinning one voice per series, and having someone other than the writer judge the script before production.
By Popcorn · Published October 26, 2026
Popcorn publishes this guide and makes AI video. The lessons come from the rules Popcorn's own production system follows. Many of those rules were written after a video failed in exactly the way they describe.
Why do viewers spot AI video so quickly?
Because they're looking for it, and they don't love what they find. In a December 2024 study by NielsenIQ, more than 2,000 people watched AI-generated ads ranging from low to high quality. They picked out most of the AI ads on their own and described them as more "annoying," "boring" and "confusing" than traditional ads. Brain measurements on about 150 of them showed weaker memory activation for the AI ads, even the ones viewers rated as high quality.
That last finding is the important one. Prettier footage alone doesn't close the gap. What viewers read as "cheap" is mostly everything around the footage: what the video says, how its shots hold together, how it sounds and how it's cut.
Is the script the real problem?
Usually, yes. The rubric Popcorn's script reviewer works from puts it this way: AI slop "is not a script with errors — it's a script that is competent, correct, continuity-clean, and dead." A flawless script about nothing in particular is the most common cheap video there is.
Five habits give it away. Count them in any script before it goes into production:
| The tell | What it sounds like | What to do instead |
|---|---|---|
| The "not X, but Y" line | "It's not a tool. It's a partner." | Say the one thing it does, plainly |
| The fake-wisdom ending | "And that's what real connection looks like." | End on a concrete image or a line only this video could say |
| Buzzwords as proof | "Seamless, intelligent, next-generation." | Show one number, one customer or one before-and-after |
| The reassurance | "And that's okay." | Commit to a position the viewer could disagree with |
| The sweeping opener | "In a world where everything is changing…" | Open on the most specific thing in the video |
Two tests catch what the list misses:
- The swap test. Replace your brand name and the one specific detail with a placeholder. If the video could now belong to any competitor, it's generic, however good it looks.
- The one-sentence test. Say what the video should make a viewer feel or realise, in one sentence. If you can only name its topic, it has no point, and no amount of editing will add one. Popcorn's reviewer rubric says it bluntly: "You can't line-edit a point into existence." Send it back to the idea.
One video should carry one idea. If two compete, make two videos.
What makes the pictures look cheap?
Video models have predictable failure modes. Each has a known fix.
| What looks cheap | Why it happens | What to do instead |
|---|---|---|
| Garbled words on signs, screens and titles | Video models draw text as shapes, so it comes out misspelled and drifting, and it can't be fixed afterwards | Keep text out of the generated scene. Put every word and number in graphics layered on top |
| A face, outfit or prop that changes between shots | Each shot is a separate generation, and a written description of a person drifts from one to the next | Build each recurring character, product and set from reference photos, and reuse the same ones in every shot |
| The same presenter, in the same frame, for a whole minute | The easiest thing to prompt, repeated | Change the setting, the framing or the action at every cut, and mix short shots with longer ones |
| Camera swoops and a glossy, oversaturated look | Models default to spectacle | Move the camera only when something motivates it |
| A collage of "cool shots" | Shots generated one by one, with no thread between them | Make each shot exist because of the one before it |
| The usual clichés: a lone figure in the rain, neon cyberpunk, a fork in the road | The first idea is everyone's first idea | Use specific behaviour and props from your own world |
| Strange hands | Models can't hold a specific finger gesture | Stage common, whole-body actions, not delicate gestures |
| Objects that appear from nowhere, vanish or duplicate | The model loses track of what is in the scene | Watch every shot for it before you publish |
Two of these have surprised even experienced teams:
- Hands don't improve with better prompting. Popcorn found that a named finger gesture, such as one finger raised, failed across every image and video model in its production stack, however explicitly it was described. Before the rule existed, a single shot could be regenerated 10 to 20 times without success. The fix is to rewrite the action, not the prompt.
- Two separate generations never match at a cut. If one shot ends on a medium shot of the host and the next begins on almost the same medium shot, the small differences read as a glitch. Change the framing at every cut, from wide to close or from one angle to another, and the difference reads as editing.
What makes the sound give it away?
Sound gives AI video away as often as the pictures do.
- The narrator's voice changes partway through. If nothing fixes the voice once for the whole video, separately generated scenes can each pick a different one. Choose one narrator voice per series and pin it.
- A character speaks in someone else's voice. With two speakers, give each a distinct voice and say so explicitly in every scene they appear in.
- The volume jumps between shots. Dialogue generated with the footage comes out at different levels from take to take. In Popcorn's own measurements it swung by up to about 3 loudness units either way between takes, enough to hear as a dip at a cut. Level all the dialogue to one target in the final mix.
- Words in capitals get spelled out. Many voice models read "AMAZING" as A-M-A-Z-I-N-G. Write scripts in sentence case and mark emphasis another way.
- The music belongs to a different film. One score written for the whole piece holds it together; a stock bed under each scene doesn't.
What makes the edit look cheap?
- Transitions everywhere. Crossfades and wipes chosen by default look like a template. Since August 2026, Popcorn's editor no longer chooses transitions on its own: every cut is a hard cut unless someone asks for a transition that means something.
- A metronome rhythm. If every shot runs five seconds, the video ticks. Vary shot lengths with what each moment needs.
- An ending that stops on the last word. Let the final image breathe for a moment after the last line.
- A soft, stretched picture. Generated shots are often made below full HD and scaled up; Popcorn's are made at 720p. Show them at the size they were made for, and don't crop in on footage that has already been upscaled.
What should you check before you publish?
Popcorn's script room uses a short ritual to judge scripts. It works on finished videos too:
- Watch it muted. With every spoken line gone, does the story still read?
- Pause on the first frame. As a thumbnail, would you stop scrolling?
- Look at the last image. Does it pay off what the opening promised?
- Say the idea in one sentence, without jargon.
- Run the swap test.
- Say honestly whether you would watch it to the end. A reviewer who flatters is worse than none; a false pass costs a whole production.
Where should a person step in?
At three points, because each one is expensive to fix later:
- The idea, before any script exists. A pointless idea can't be rescued by craft.
- The script, judged by someone other than the writer. Popcorn separates the two on purpose: one agent writes, and a separate reviewer agent judges, with a standing bias against safe, bland drafts. In its rubric, "a timid, safe-in-every-direction draft must score worse than a bold draft with a rough edge."
- The final watch-through, before anything is posted.
Then decide shot by shot what should be real. The people, products and moments that have to be true belong on real footage; generated scenes work best around them, not in their place.
How does Popcorn handle these?
| The cheap tell | Popcorn's approach |
|---|---|
| A script with no point | A writer agent drafts, and a separate reviewer agent challenges each script for generic storytelling before production. The plan waits for your approval first |
| Garbled text | Titles, lower-thirds, stat cards and charts are motion graphics built in code, so every word and number comes out exactly as written |
| Faces and props that drift | Recurring characters, products, sets and voices are built once and reused; characters can be built from photos of real people you have the rights to |
| Fake-looking people and products | Your own footage is placed exactly as filmed, never regenerated |
| A video that sounds like everyone else's | A brand bible from interviews with your team and research into your market, attached to every brief |
Popcorn does not catch every problem in this guide automatically. Its videos still need the final watch-through above, and its edits can be changed shot by shot in chat or on the timeline.
For keeping a series consistent across episodes, see how to turn a blog, newsletter or webinar into a video series.
Where do these facts come from?
- Viewer reactions to AI-generated ads: NielsenIQ, NIQ research uncovers hidden consumer attitudes toward AI-generated ads, December 12, 2024, checked 2026-10-26.
- The script tells, tests and checklist: the script-review rubric and production rules inside Popcorn, as of October 2026.
- What Popcorn does: the Popcorn FAQ, including brand consistency, real footage, your own face and voice and resolution.
Frequently asked questions
Why does AI video look fake?
Usually because of the parts around the footage: generic scripts, garbled text inside generated scenes, characters whose faces or clothes change between shots, repetitive framing, voices and volume that shift between shots, and default transitions.
Can you put text inside an AI-generated video?
Not reliably inside the generated scene: video models draw text as shapes, so it comes out misspelled and cannot be corrected afterwards. Put titles, numbers and captions in graphics layered on top of the footage instead.
How do you keep an AI character consistent from shot to shot?
Build the character from reference photos rather than a written description, reuse the same character in every shot, and change the framing at each cut so small differences between separate generations read as editing rather than glitches.
Do viewers dislike AI-generated ads?
Often. In a December 2024 NielsenIQ study of more than 2,000 people, viewers identified most AI-generated ads on their own and found them more annoying, boring and confusing than traditional ads, and even ads rated high quality produced weaker memory activation.
What is the single biggest fix for cheap-looking AI video?
Fix the idea before the footage. A video needs one point you can state in a sentence and details that belong only to your brand; if swapping in a competitor's name would leave it unchanged, no amount of visual polish will save it.