Not a real recording.
It says so first.
Michael Lip, MakeRoll's founder, made an AI version of himself from one terrace photo and a 2025 recording of his voice. HeyGen made the avatar. Fish Audio cloned the voice. A cheap check sat in front of every paid step, and the checks caught six defects before the final cut.
Disclosure rules: MakeRoll's AI video rights checklist ↗. The film follows its platform AI-label rules. It is a labelled promo. It is never a funding video or a testimonial.
01 / WATCH THE FILM
It says it is AI.
In the first second.
The first spoken line is "This is an AI version of me." An AI label sits on every frame. The end card names the tools and says he gave consent.
26 seconds. A label on every frame.
The avatar opens and closes the film. In between, MakeRoll test ads with AI characters play under the cloned voice, with a green tag that says so. The 9:16 cut is the main version. It was cropped from the 16:9 render.
Avatar by HeyGen. Voice clone by Fish Audio. The B-roll is MakeRoll test ads with AI characters. The platform label rules we followed are in our rights checklist. Open the canonical reference post ↗
Read the visual transcript and end-card copy
Every spoken line is an AI voice clone of Michael Lip. 00:00 to 00:05. AI avatar on a terrace. Labels: "AI avatar of Michael Lip · AI-generated" and "Not a real recording. AI avatar + AI voice clone of Michael". Voice: "This is an AI version of me. I never said these words. I'm Michael, and I built MakeRoll." 00:05 to 00:11. MakeRoll test ads with AI characters, tagged as such. Voice: "One brief goes in. 5 to 10 AI agents cast, render, voice and cut the ad." 00:11 to 00:15. Avatar. Voice: "A gate with 25 checks scores every ad. In our first full run," 00:15 to 00:22. MakeRoll test ads with AI characters. Voice: "3 out of 4 ads scored 95 or more on our own 25-check gate." 00:22 to 00:24. Avatar. Voice: "The link is on screen." On screen: makeroll.com/demo. 00:24 to 00:26. End card: MakeRoll. makeroll.com/demo. AI avatar of Michael Lip, made with his consent. Avatar: HeyGen · Voice: Fish Audio clone · not made with MakeRoll.
02 / HOW IT WAS BUILT
One photo.
One old recording.
The avatar comes from one terrace photo. The voice is a clone built from a 2025 recording of Michael. HeyGen and Fish Audio made the film. MakeRoll-style checks decided what got through.
The only image the avatar saw.
The real photo, cropped and nothing more. We tried an AI clean-up first and threw it out, because it redrew his face. Identity against his real photo fell to 0.826.
The wide grin in the photo carried into the film. Reviewers call it a constant grin. A new closed-mouth photo is the next fix.
Clone the voice
Fish Audio S2.1 cloned Michael's voice from 198 seconds of a 2025 recording. An older 2024 source reached only 0.67 similarity, so we dropped it.
Score every take
44 promo takes were scored against real audio the clone never heard. The winner reached 0.787. Then speech recognition checked the words, and takes heard as "Make a Roll" or "MakeRow" were rejected.
Preview before the full render
HeyGen Avatar IV rendered a 5-second preview first. The vertical preview left the lips open on b, m and p, so the full vertical render never ran. The 9:16 cut was cropped from the 16:9 render instead.
Cut, label, measure
Pauses were cut, B-roll was added and every frame got an AI label. The build fails if one frame lacks a label. A review found 190 to 280 ms of A/V drift. The fix measured 0 ms on the final file.
Read the check log and every review score
The six defects the checks caught
1 Voice best-sounding take said "Make a Roll" rejected 2 Voice every first-script take heard as "3 or 4" line rewritten to "out of" 3 Voice a take heard as "MakeRow", a live site rejected 4 Avatar vertical preview: lips open on b, m and p full render blocked 5 Edit A/V drift 190 to 280 ms (blind review) fixed, 0 ms on the final file 6 Edit spliced audio 60 to 180 ms short (build) fixed in its own pass
Blind review scores out of 100
version creative 16:9 / 9:16 trust 16:9 / 9:16 v2 56 / 56 69 / 65 v3 62 / 58 75 / 72 v3.1 not scored 74 / 72 v3.2 63 / 57 79 / 72 v3.4 68 / 64 80 / 78 v3.5 not re-scored v3.6 not re-scored (cost line cut, 26 s)
Fresh reviewers each round, so no one anchored on an earlier version. One lens judged the film as an ad. The other judged honesty and disclosure.
03 / TECHNICAL REVIEW
Trust scored well.
Creative did not.
Two fresh blind reviewers scored each version. One judged it as an ad. One judged honesty and disclosure. Read both numbers together.
best creative score, v3.4 16:9
The trust lens gave the same cut 80. The creative lens says the face and voice still read as AI and the call to action is weak. The 9:16 cut scored 64 and 78. v3.5 fixed the next defects but has not been scored again.
04 / PRODUCTION DETAILS
The full breakdown
is for members and partners.
Production charges, metered usage and the per-call ledger are shared in the MakeRoll member room or with business partners. The public case keeps the film, method and review available to inspect.
Access production details
05 / WHAT THE EVIDENCE DOES AND DOES NOT SAY
A labelled promo.
Not proof of anything else.
One careful build is a data point. Here is what this case can and can't support.
No audience result yet.
This page claims no reach, clicks or sign-ups. It records how the film was made and checked.
It still looks and sounds like AI.
The best creative score was 68 out of 100. The brand name sounds like "Makro", the grin never stops and the newest cut has no fresh score.
Never for funding forms or testimonials.
Some funding programs ban AI founder videos in writing. Michael records those on camera. This film is a labelled promo and nothing more.
Tools and consent
Avatar: HeyGen Avatar IV. Voice: Fish Audio S2.1 clone. Both use Michael's own photo and voice, with his consent. The B-roll is MakeRoll test ads with AI characters. The platform label rules are in our AI video rights checklist ↗.
06 / QUESTIONS
What the numbers
can support.
- Is this a real recording of Michael?
- No. Michael never said these words on camera. HeyGen animated one photo, and a Fish Audio clone of his voice reads the script. The film says so in its first line.
- Was it made with MakeRoll?
- No. HeyGen and Fish Audio made it. We ran MakeRoll-style checks on it: short previews, speech checks and blind reviews.
- Did it pass every check?
- No. The blink rate is low, the brand name sounds like "Makro" and the final cut has no fresh review score. The last scored cut got 80 for trust and 68 as an ad.
- Can it go in a funding application?
- No. Some programs ban AI founder videos in writing, and others ask for the founder on camera. Funding videos are real takes. This one is only a labelled promo.
- Why is the face score higher than his own photos?
- The 0.92 compares the avatar with the one photo it came from. His real photos match each other at only 0.60 to 0.76. So the score shows the avatar stays close to that photo, not that it looks more like him than he does.
- How close is the voice?
- 0.787 similarity to real audio the clone never heard, the best of 44 takes. That is a speaker-match score from a model, not a listener test.
ONE BRIEF AT A TIME
Check every take.
Keep the label on.
MakeRoll is a weekly AI video build room. Work from a brief, check each take, label what is AI, and keep the limits visible.
See membership options