UGC Handbook / Lesson 07 of 14 / Chapter II: Make

AI-generated UGC: when to use it, when to avoid it

By Mojo's math, testing five human creators takes $2,500 and three weeks. Testing five AI variations takes $100 and two hours.

And most of the AI variations failed on contact.

Both facts come from the same RevenueCat guest post from the team at Mojo, a video-editing app, after they tested AI-made user-generated content (UGC) ads against their own human-made ones. The point of this lesson is not "AI good" or "AI bad." It's that AI does one job extremely well and another job badly, and most founders ask it to do the wrong one.

Where AI UGC fails: inventing a person

Mojo's best-performing ad was a plain 30-second split-screen: their Product Manager explaining the product on top, a screen recording of the app on the bottom. It converted trial users to paid at 23%. Their read on why: he built the feature, and "his conviction came from ownership."

Encouraged by an earlier win (below), they moved to fully AI-generated avatars. The post titles this phase "the 80% failure rate," and the body is blunter: "most of the profiles we tested failed almost immediately."

Why they failed had little to do with pixels. Three reasons, in Mojo's words:

• Staging. The generic avatars sat in hyper-polished environments: studio microphones, cinematic lighting, stylized backgrounds, some at a three-quarter angle instead of looking into the lens. "The user's brain categorized the content as a commercial within the first second."
• Casting. The avatars looked like stock models or polished influencers, not like the raw, founder-led person whose credibility they were trying to scale.
• Tiny sync errors. One avatar had a 0.2-second lip-sync delay on a key word. Click-through rate dropped 68% compared with the human baseline. "Users couldn't articulate why they did not trust it, but the data was brutal."

Formats that run on trust reject synthetic faces. In Mojo's tests, fully synthetic avatars "struggled in formats that rely on credibility, especially testimonials or personal stories." Their rule: "use AI to explain, and use humans to convince."

And some markets reject them outright. The same dubbed avatar that worked in Brazil and Spanish-speaking markets failed in France and Germany. Mojo's conclusion: sensitivity to synthetic media "is geographic and culturally dependent. You can't assume universal adoption."

Joseph Choi of Consumer Club made a related prediction on the Superwall podcast in 2025. Right now, he said, AI avatars get treated as normal people in the comments (his example: viewers dutifully commenting the call to action (CTA) word under an AI-generated face).

But "once people catch on that anyone can just fake a face," he expects trust to shift back toward voices. He also flagged the ethical concerns and left them to the viewer's judgment.

Where AI UGC works: replicating a human winner

Now the earlier win, which Lesson 01 told in full. When local creators re-shot the PM's ad for other markets, performance dropped; when Mojo instead had HeyGen dub the original while preserving its timing, pauses, and gestures, the Brazilian version acquired customers at 40% lower cost than the native creators. Mojo's lesson: "AI does not reinterpret; it preserves."

The one avatar that worked was a copy, not an invention. Instead of a generic profile, they built a HeyGen digital twin strictly of their own PM. They stripped the studio mics and cinematic lighting and matched his raw facecam angle, look, and baseline energy.

That version reached 87% of the original human conversion rate. Cost: $20 to generate versus $500 for a traditional creator. Result: 31% lower cost per acquisition.

Then they stacked it. Custom avatar for the base video, ElevenLabs voice cloning to localize it. Brazil numbers from that "double AI stack": cost per acquisition $8 (31% below the human control), click-through 4.5%, conversion 3.1%, return on ad spend 2.1x. And again: it broke in Europe.

Look at the pattern across all three wins. The source was always a real human who had already converted. AI changed the language, the voice, or the delivery medium, never the person or the performance.

Break the perfection. Mojo's most practical tip: AI "naturally gravitates toward a flawless output, so you have to actively force it to be messy." Lower the resolution slightly, and keep some background room tone in the audio. Perfect delivery reads as a commercial.

Peter, co-founder of the fitness app Stronger Now, made the mirror-image point on Superwall's channel: most vibe-coded apps go faceless now, so being willing to put a real face on camera is itself an edge.

The new bottleneck is judgment, not production

When you can make 50 variations a day, intuition fails. That's Mojo's phrase. With five videos a month you can pick winners by gut. With 50 a day you need a decision rule, and "bad decisions now scale just as fast as good ones."

Rights. If your growth engine runs on a digital copy of a person, you need to own that copy. Mojo's warning: with external creators, "if you don't secure rights to use their digital likeness, you're essentially renting your growth." They recommend likeness agreements that spell out how, where, and for how long a synthetic version can run, including a clause covering six to twelve months after an employee leaves.

The trust tax. For a subscription app, a user who feels tricked by an ad "might install the app, but will inevitably churn." Mojo's term for this is a trust tax, and it lands on retention, which is where your revenue actually lives.

So decide your boundary before you generate anything:

Do this before you open an AI video tool

1. Have a human winner first. If no human-made video has produced installs or trials yet, you have nothing worth replicating. Go back to Lesson 05.
2. Check the format, using Mojo's own split. Walkthrough, feature demo, or screen-recording tutorial: AI is allowed. Testimonial, storytime, or founder intro: humans only.
3. Confirm you own the likeness. Written rights to the face and voice you're cloning, with a duration. No agreement, no clone.
4. Copy, don't cast. Build the avatar from your actual winner's face, framing, and energy. Never pick a "better-looking" generic profile.
5. Match the mess. Same raw facecam angle, slightly lowered resolution, room tone left in. If it looks like a studio, start over.
6. Test one market you already know converts before touching a new one, then test tolerance market by market.
7. Set the kill rule up front. Mojo's checklist question: can you kill the campaign in 24 hours if the data says it's failing? Write the threshold down now.
8. Judge on subscription events, not clicks. A variant that matches the original's click-through but loses on trials is a failed variant.