The Duolingo output reads as a casual creator-style presenter rather than an AI-obvious stock avatar, with the report calling the setting genuinely UGC-authentic.
What was measured
Avatar realism
Believable presenter; accurate lip-sync; how uncanny or AI-obvious the presenter looks.
decisive for this rankingtransformation
If the presenter looks fake or lip-sync is off, the ad fails at the core UGC avatar job. (3 of 3 judges)
What was given, what came back
Test input: Duolingo – Mobile App · mixed · group: ugc-style-video-ads
Input — what we sent
Input, verbatim
Product URL: https://www.duolingo.com/ Exact script: Hook: Learning a new language has never been this easy. Script: I've been using Duolingo for a few minutes every day, and it's made language learning simple and fun. The short lessons are easy to follow, and the daily practice keeps me motivated. If you're planning to learn a new language, give Duolingo a try. Output requirements: - Vertical (9:16) - 20–30 seconds - UGC-style AI presenter - Natural AI voice - Captions enabled
A short-form UGC promo prompt for a consumer mobile app, designed to test whether the tool can explain app benefits simply and persuasively while generating a realistic AI-presenter video.
Why this input is hard
- · Mobile app understanding
- · Consumer app promotion
- · Script quality
- · Avatar realism
- · Voice quality
- · Caption quality
- · Overall ad readiness
Output — unretouched
Also checked on this input — same tool, 5 other criteria
B-roll / product integration✗ FailedThe supporting footage is not product-integrated because the cutaway uses the wrong app entirely, so the B-roll does not match the Duolingo testimonial being spoken.Caption quality◐ MixedCaptions are present, but around 0:04-0:05 they overlap with a leftover badge graphic and become garbled, creating hard-to-read double-exposed text.Platform formats✓ WorkedThis is the only run that delivers the requested full-bleed vertical 9:16 frame with no letterboxing.Product understanding✗ FailedProduct understanding breaks in the B-roll cutaway: the phone screen shows a shopping app ('NEW YORK BAG $219, ADD TO CART') instead of Duolingo or any language-learning interface.Script quality & persuasiveness✓ WorkedThe script is delivered verbatim from start to finish and lands on a natural CTA, ending correctly on 'give Duolingo a try.'
Provenance
- Observation
- 9012a876-7dac-4276-89c3-aa9dc9ebcf9d
- Evidence run
- 9fdbb84b-4aa0-484d-8f89-5c46608019a5
- Study
- Generate UGC-Style Video Ads With AI Avatars
- Research task
- 86baxegpg
- Tested at
- not recorded
- Source
- first-party
- Evidence state
- verified
- Proof shown
- input + output shown
- Cost / latency
- not captured
- Repeat run
- not captured
- Tester
- not captured
The last three rows are honest blanks, not placeholders — our capture has no field for them yet.
Query this
get_evidence({
tool: "heygen",
scenario: "ugc-style-video-ads"
})MCP · mcp.aidemos.com/api/mcp
Free with attribution.
Same input, same check — 7 other tools
measured on Avatar realism
AKOOL◐ MixedThe presenter is serviceable, but the avatar still looks moderately AI-generated rather than high-end realistic.DreamFace◐ MixedThe Duolingo avatar has good facial realism and natural gestures/weight shifting, but the right hand turns into a smooth mitten-like blob around 0:11.Invideo AI✓ WorkedThe presenter reads as a consistent human avatar with a clean UGC look throughout the clip, with no obvious flicker or identity break reported.JoggAI✓ WorkedGenerates a strong-looking photorealistic avatar with no visible artifacts; the report says the face, hair, and teeth look natural even in a wide smile.Synthesia✓ WorkedThe Duolingo output used a smooth, centered talking-head avatar, and the report says lip-sync accuracy was maintained.Topview AI◐ MixedThe avatar remains visible and moves slightly, but the report describes the presenter as more AI-like than premium tools.Vidnoz AI◐ MixedProduces a believable talking-head presenter, but the report still describes the avatar as moderately AI-generated compared with higher-end tools.
This evidence is published in
Real inputs and real outputs, no retouching · every cell queryable via API & MCP · aidemos.com