A synthetic person delivering your script is the most recognisable form of AI advertising and the most contested. It works well for some jobs and poorly for others, and the line between them is sharper than the marketing suggests.
What it is
A generated human — not a real actor, not a cartoon — speaking lines you wrote, usually to camera, usually in a vertical frame. The face is synthesised, the voice is synthesised, and the lip movement is matched to the audio.
The quality range is wide. At the good end, a casual viewer scrolling past does not register anything unusual. At the bad end, the uncanny valley is immediate: eyes that do not quite track, a mouth that moves correctly but a face that does not, an emotional flatness underneath natural-sounding words.
Where any specific tool falls on that range changes every few months, so judge output rather than claims.
What it is genuinely good for
Explaining something. When the selling point is a fact rather than a sight — a guarantee, a compatibility, a comparison — someone saying it is more efficient than showing it. This is the strongest case.
Volume across a catalogue. Forty products each needing a thirty-second explainer is not a shoot anyone will fund. It is a reasonable use of generation.
Testing scripts before committing. Before paying a creator to deliver a script, finding out which of five scripts works is cheap this way.
Languages you cannot cast for. Generating the same explainer in six languages is plausible; casting six presenters is not.
Formats where the presenter is incidental. A voice over product footage with brief presenter cutaways carries far less uncanny-valley risk than a full talking head, and often performs better anyway.
Where it fails
As social proof. A synthetic person saying they use and love the product is an advertisement wearing the costume of a testimonial. Audiences increasingly read this, and the cost of being caught is trust rather than performance. Do not put generated presenters in a reviews module and do not write scripts in the first person about personal experience that did not occur.
When the product needs handling. Hands remain the weak point of generative video. A presenter demonstrating a product is exactly the shot most likely to fall apart.
For high-consideration purchases. The more money is at stake, the more scrutiny a face receives, and the more a slight wrongness costs.
With audiences who are saturated. Some niches have seen so many AI presenters that the format itself is a negative signal. This varies enormously and is worth testing rather than assuming.
Long-form. Attention to detail increases with duration. Fifteen seconds forgives what ninety seconds does not.
Disclosure, which is tightening
This section will date faster than the rest, so verify current rules before building a campaign around them.
The direction of travel is clear and one-way:
- Platforms increasingly require labelling of realistic AI-generated content, and increasingly detect it automatically whether or not you declare it.
- Advertising regulators in several jurisdictions have moved against synthetic testimonials specifically — a generated person presented as a real customer is treated as a false endorsement, which is a different and more serious category than an undisclosed ad.
- Likeness rights matter if the presenter resembles a real person. Several jurisdictions have introduced or strengthened rules here.
Three practical rules that survive most rule changes:
- Never present a synthetic person as a real customer. This is the line that attracts actual penalties.
- Label it when the platform asks. Undisclosed-then-detected is worse than disclosed.
- Keep claims truthful regardless of who speaks them. A synthetic mouth does not launder an unsupported claim.
Writing a script that does not sound generated
Most AI presenter ads are undermined by the writing rather than the rendering.
Write how people talk, not how copy reads. Contractions, short sentences, the occasional fragment. "It's about the size of a credit card" beats "It measures approximately 85mm in width."
Open with the problem, not the product. Ten words. "My kitchen drawer was a disaster" earns the next five seconds.
One claim. A thirty-second video supporting one specific claim outperforms one listing six features. Specificity is what survives.
Steal from your reviews. The phrasing customers already use is pre-validated and sounds like a person because it was one.
Read it aloud and time it. Roughly 2.5 words per second for natural delivery. A fifteen-second video is about 35–40 words. Most first drafts are twice that, and overlong scripts are the main cause of rushed, synthetic-sounding delivery.
End with one action. Not three.
Making it land
- Cast against the stereotype. The polished studio presenter reads as an ad. Someone who looks like the customer reads as a recommendation.
- Put the presenter somewhere real. A kitchen, a car, a bedroom. Plain backdrops signal production.
- Show the product in cutaway, not in hand. Sidesteps the hands problem and keeps attention on the item.
- Always caption. Most of the feed is muted; it also hides small lip-sync imperfections.
- Keep it short. Fifteen seconds forgives much.
- Match age and appearance to the audience honestly. If the product is for children, the presenter should not be a corporate adult — and if the product is used by a child, think carefully about whether a synthetic child is appropriate at all.
A reasonable testing approach
- Three scripts, one presenter. Find the script first; it matters more.
- Then three presenters, winning script. Fit between presenter and audience is real and not predictable from the armchair.
- Compare against a faceless version of the same claim. Often the faceless version wins, and knowing that is worth the test.
- Watch hook rate specifically. A presenter that triggers the uncanny-valley response shows up as a collapsed three-second rate.
That last test is the one most people skip. The useful question is not "does my AI presenter ad work" but "does it beat the same ad without a presenter" — and sometimes it does not.
Where Faysell sits
Faysell generates ads with AI actors from a product photo, carrying the character's appearance across every shot so the presenter does not change between cuts. The AI actor page covers that path; the faceless path covers the alternative worth testing against it.
Honestly: full lip-synced delivery is the area where generative video remains hardest, and presenter-led output is better treated as one format to test rather than the default. For many e-commerce products the faceless version performs at least as well and carries none of the disclosure complexity.
The summary
AI spokespeople are good at explaining and at scale, bad at social proof and at handling products. Never present one as a real customer — that is the line with actual legal weight.
Write short, write spoken, open with the problem, support one claim. And always test the faceless version alongside, because it wins more often than the category's marketing would suggest.