What makes AI UGC look fake: 8 tells and the fix for each
Viewers rarely spot the face. They spot the edit around it. The eight tells that give synthetic creators away in a feed, ranked by how often they show up, with the production fix for each one.
Creative6 min read
Key takeaways
- The face is usually fine at feed size. The pacing, the hands, the edit, and the script are what give an AI creator away.
- Most tells are fixed upstream: a better reference clip, a better portrait, or a shorter take.
- Cut away often. A synthetic performer in two-second stretches over real footage is nearly invisible. The same performer holding for thirty seconds is a tell in itself.
- The script is the biggest tell of all. If it reads like a product page, no render will save it.
Start by watching like a viewer
Teams evaluate AI creators by pausing a render at full screen and inspecting the teeth. Viewers evaluate them at phone size, mid-scroll, with the sound off, for two seconds. Those are different tests, and the second one is the only one that matters. Most of what fails the first test passes the second. Most of what fails the second test has nothing to do with the face.
So before you change models or tools, run the viewer test. Put the ad in a feed with three real posts around it and watch it on a phone. What you notice in the first two seconds is the tell. It is almost always one of the eight below.
Tell 1: even pacing with no breath
Real people speed up, slow down, pause to think, and breathe between clauses. Early avatar tools delivered every syllable at the same rate, and the result read as a text-to-speech voice with a face. It is the most common tell and the most fixable.
Fix: use a motion reference, not a script alone. Reference-driven renders copy the timing of a real human performance from the reference clip, including the pauses. Choose a reference where the person is talking naturally, and trim it to the segment whose rhythm you want. If your tool only takes a script, write the pauses in and shorten the sentences.
Tell 2: dead hands
Hands that never move, or that float in the same position for ten seconds, break the illusion faster than any facial artifact. People talk with their hands, especially when they are showing you something.
Fix: pick a reference clip with moderate, natural hand movement, and keep the take short so the hands do not have time to look repetitive. Cut to a real hands-and-product cutaway whenever the script mentions the product; that is where the viewer expects hands anyway.
Tell 3: the thirty-second single shot
A real creator filming on a phone would never hold one framing for thirty seconds. They would jump-cut, zoom, turn the phone to show something. A synthetic performer that holds the frame the whole way gives the viewer thirty seconds to find something wrong, and they will.
Fix: cut away every two to three seconds. Six to ten B-roll cutaways in a thirty-second ad is the norm for human UGC, and it does double duty for AI creators by keeping each stretch of synthetic performance short. We wrote a full cutaway plan.
Tell 4: studio light on a bedroom wall
Skin that is perfectly even, catchlights that are too clean, and a background that is suspiciously tidy all say "generated" before the viewer can articulate why. Phone footage is a little noisy, a little uneven, and lit by whatever was in the room.
Fix: start from a portrait that was actually shot on a phone in a normal room, not a studio headshot. Match the B-roll's color and grain to the talking head. If your tool offers a film or phone look, use it lightly. Perfect is the tell.
Tell 5: captions that do not match the mouth
Auto-captions that run early or late, or that paraphrase the audio, draw the eye to the mouth, which is where a synthetic performer is weakest. Mismatched captions also break the muted-viewer experience, which on Meta placements is most viewers.
Fix: caption from the final audio, word for word, with timing checked by eye. Use a caption style that reads at thumbnail size and sits away from the mouth. If the script changed after the render, re-render or re-caption; never leave the two out of sync.
Tell 6: the script sounds like a product page
This is the biggest tell and the only one no render can fix. "Introducing our revolutionary formula, clinically designed to..." is not how anyone talks to a phone. Viewers do not consciously identify the AI; they identify the ad, and they scroll.
- Write for the ear. Read it aloud; if you would not say it to a friend, cut it.
- One idea per sentence, and short sentences. Feed-native speech is choppy.
- Lead with the viewer's problem, not the brand's adjectives.
- Cut every superlative. Specific beats revolutionary.
- Keep first-person experience claims out of a synthetic performer's mouth. That is both a tell and a compliance problem.
Tell 7: identity that drifts between cuts
When an ad cuts between two renders of the "same" character and the face is subtly different, viewers feel it even if they cannot say what changed. It reads as uncanny rather than fake, which is worse.
Fix: one saved portrait per character, used for every render, and a portrait that follows the rules in the cast guide: front-facing, evenly lit, uncovered, neutral. Review the first render against the portrait before you cut variations.
Tell 8: a reaction that does not match the line
A big smile on a sentence about a problem, or a flat face on a reveal, is a timing mismatch between the reference's emotion and the script's content. Humans do this too when reading a teleprompter badly, and it looks just as wrong on them.
Fix: choose the reference clip for its emotional shape, not just its duration. If the script goes problem, turn, proof, the reference should go frustrated, curious, pleased. When you cannot find one reference with the whole arc, split the ad into two short renders and cut between them with B-roll.
A quick audit before you publish
| Check | Pass if |
|---|---|
| Pacing | There is at least one natural pause and the speed varies |
| Hands | They move, and real hands appear in cutaways when the product is mentioned |
| Shot length | No single stretch of synthetic performer exceeds about four seconds |
| Look | Talking head and B-roll match in color, grain, and camera motion |
| Captions | Word-for-word, timed, readable at thumbnail size, away from the mouth |
| Script | You would say it to a friend; no product-page adjectives; no first-person experience claims |
| Identity | Every cut of the character comes from the same saved portrait |
| Emotion | The face matches the line at each beat |
| Disclosure | AI presenter disclosed on screen where required |
Timing from a real performance
Copy Video renders your character with the pacing of a reference clip you own, so the pauses and gestures come from a person.
Frequently asked questions
- Can viewers tell an ad uses an AI creator?
- At feed size, in short stretches, usually not from the face. They notice even pacing, dead hands, long single shots, mismatched captions, and product-page scripts. Fix those and most viewers will not consciously register the performer as synthetic. Disclose it anyway where the rules require.
- Which AI video model looks most realistic?
- Model choice matters less than the reference, the portrait, and the edit. A mid-tier model with a natural motion reference and frequent cutaways beats a top-tier model holding a thirty-second single shot.
- Should I hide that the creator is AI?
- No. Several jurisdictions require disclosure of synthetic performers in ads, and platforms label AI content. Make the ad good enough that the disclosure does not matter, rather than trying to avoid it.