"AI commercials are terrible" is a fair verdict on most of them, and the reasons are predictable: uncanny people, physics that melts, products that drift between shots, prompts that all produce the same look, and no actual idea underneath. None of those is a limit you must live with. They are choices.
Why AI commercials look bad: five failure modes
1. Uncanny people. Faces, hands and the way bodies move are where viewers are most sensitive. When Coca-Cola remade its "Holidays Are Coming" spot with AI in 2024, critics singled out the human figures and "challenges in creating lifelike human expressions and movements," while the animal-focused spots in the same campaign drew less criticism (Social Samosa).
2. Drift. The same character or product looks different from shot to shot. A self-described AI videomaker reviewing the 2024 Toys"R"Us Sora film flagged hands and keeping a character consistent from scene to scene as the weak points (Decrypt). For brands, drift usually shows up as a label that gains letters or a logo that changes shape.
3. Mush physics. Liquids that pour upward, fabric that passes through itself, wheels that slide instead of turn, objects that merge on contact. Every second is another chance to break a rule the viewer knows by heart.
4. The default look. Golden hour, shallow depth of field, slow push-in, teal shadows. Prompt "cinematic" and you get the same average every brand got.
5. No idea. A montage of pretty shots is not a commercial. Most bad AI ads skipped the step where someone wrote down, in one sentence, what the ad is about.
The AI ad backlash, in numbers
In the IAB's January 2026 study, 82% of ad executives believed Gen Z and millennial consumers felt positive about AI-generated ads, while only 45% of those consumers actually did (IAB). Among Gen Z, 30% described brands using AI for ads as "inauthentic."
The same study has the more useful half of the story: 73% of Gen Z and millennial consumers said knowing an ad was made with AI would either increase their purchase likelihood or make no difference, and high-quality visuals ranked first among what drives their attention. People are not rejecting AI. They are rejecting bad work.
And it is not cheap to do well. The Kalshi ad, often cited as proof AI ads can be fast, took "about 300–400 generations to get 15 usable clips," and its maker told NPR: "Just because this was cheap doesn't mean anyone can do it" (WUNC/NPR). See what AI video ads cost for where that iteration lands in a budget.
How to make an AI ad that doesn't look bad
- Concept first. Write the idea in one sentence and sketch six frames. If the sketch is boring, no model will save it.
- Stylize on purpose. Pick a look models render well and commit to it: liquid chrome, glass, macro surfaces, graphic color fields. Stylization also gives you permission to be impossible.
- Objects over people. Products have no uncanny valley. If the story needs humans, use silhouettes, hands at a distance, backs of heads, or shoot real people and keep AI for the world around them.
- Lock the product. Composite the real packshot into your keyframes and animate with start and end frames, so the model has less freedom to redraw your label. Our step-by-step for AI product video from a single image shows the chain.
- Edit like an editor. Clips are usually 5 or 10 seconds; the usable part is often the first second or two before things start to wobble. Cut before the model fails. Fast, motivated cuts hide more than another regeneration does.
- Design the sound. A glass tap, a breath, real room tone. Stock music over silent AI footage is the single fastest way to feel cheap.
- Be honest. Label realistic AI content. Hiding it turns every viewer into a fact-checker hunting for the sixth finger.
Bad vs good: the same brief, two prompts
| Generic | Directed | |
|---|---|---|
| Prompt | "Cinematic shot of a luxury watch, epic lighting, 4K" | "Watch face-up on wet black slate, single coral light raking from left, camera locked, only the second hand moves, slow" |
| Result | Default golden glow, random camera drift, hands warp | One readable image, minimal motion, nothing for the model to break |
| Why | Asks the model to decide everything | Decides everything the model is bad at deciding |
The directed prompt is not more poetic, just more restrictive. Constrain motion, light and what may change.
When AI is the wrong tool
If the shot depends on exact mechanics (a hinge, a zipper, a pour that must match a real product), generative video will fight you. That is a job for controlled 3D, which we compare in CGI product ads. If the ad depends on a human performance, cast a human.
