AI-generated video ads have a ceiling, and practitioners are already drawing the line
Fully AI-generated video ads are visibly low quality, and the practitioners running real accounts are setting a firm ceiling on how much of the creative mix they should occupy. The argument is not aesthetic. It is structural, rooted in how modern ad platforms actually read creative.
Matt Swulinski, a performance advertising practitioner, puts the quality problem in plain terms: fully AI-generated videos are, in his word, slop. Any viewer, he argues, can tell immediately that the video is AI-made. His recommended ceiling is roughly 5 percent of an account’s creative mix, used to add variety, not to carry the account. Beyond that threshold, he sees no credible path to replacing real people with AI output, at least not at present.
That view from the advertising execution side finds a parallel in live commerce. Grant LaFontaine, who leads Whatnot, says the platform currently sees no need for AI avatars because that is not why people come to the platform. The logic is product-specific but the underlying principle is shared: where human connection is the value being sold, an AI-generated substitute does not hold.
The timing of these assessments matters, and the reason has to do with how the major ad platforms have changed. Swulinski describes a Meta update called Andromeda that altered how the algorithm works at a fundamental level. Rather than letting advertisers specify audience parameters directly, Meta now analyzes the creative itself to find the right people. The creative, in other words, has become the targeting mechanism. Under that model, quality and authenticity are not just brand concerns. They carry the targeting load. A video that reads immediately as AI-generated sends a degraded signal to the algorithm before the ad reaches anyone at all.
There's an update in meta that was called Andromeda that essentially changed the targeting algorithm where the creative is the targeting. So what meta did is stop telling me in audience settings in the campaigns who people are. we're going to analyze the creative and based on who we know you're trying to target, we will find those people for you. Matt Swulinski
This is the structural argument beneath the aesthetic one. Swulinski’s ceiling of roughly 5 percent is not primarily a quality judgment about how AI video looks. It is a claim about what the creative stack can bear given how distribution platforms now work. If the algorithm is reading the creative to decide who sees it, then flooding an account with visibly artificial material does not just risk brand perception. It risks misrouting the entire campaign.
LaFontaine’s position at Whatnot sharpens the point from a different angle. Live commerce is a category where the transaction depends on a viewer’s trust in a specific human seller. The question of whether AI avatars could replicate that dynamic is, by his account, not a live question at Whatnot. The platform does not see a need for them because the human presence is not decoration. It is the product.
None of this is an argument against AI in the creative process more broadly. Swulinski’s figure is not zero. The operational advantages of generative video tools are real: speed, cost reduction, and the ability to produce variants at a scale traditional production cannot match. For creative work that does not depend on emotional authenticity or immediate human recognition, those advantages hold. The error is treating operational efficiency as a proxy for creative effectiveness, or assuming that what works for low-stakes inventory will scale upward without limit.
What Swulinski and LaFontaine are describing is a ceiling, not a floor. AI-generated video occupies a specific and bounded role in the creative stack: variety generation, rapid testing, lower-stakes inventory where authenticity is not load-bearing. At the top of the funnel, in brand work that depends on trust, in live commerce where the seller’s humanity is the draw, and in any context where a viewer’s immediate recognition of something artificial undercuts the message, human presence is not a legacy preference. It is the mechanism. The practitioners drawing these lines are not resisting AI. They are mapping where it actually works, and where the ceiling sits closer to the ground than the current enthusiasm for generative tools would suggest.