There are two AI stories in advertising and they get told as one. The first is about production economics, and it is dramatic. The second is about performance, and it is modest. Part six of the ad creative series.
Adoption went vertical
Meta reported more than a million advertisers using its generative AI ad tools, producing 15 million ads in a single month, by September 2024. Google reported roughly 70 million creative assets generated through its systems in one quarter, three times the prior year's rate. Klarna reported saving about $10 million a year — $6 million on image production and $4 million on external vendors — and cutting its image production cycle from six weeks to seven days.
Those are real, large, and mostly about cost rather than effect.
The lift did not
Here is the number that gets left out. On Meta's own site, the formal figures for its generative creative tools are an 11% improvement in click-through rate and a 7.6% improvement in conversion rate.
Set those four numbers beside each other and the honest reading is clear: the technology has transformed how much creative gets made and barely moved how well it works. An 11% click-through improvement is a good result for a production tool. It is not a different category of advertising.
This is also where a widely repeated statistic falls over. "Creative diversification delivers 32% efficiency" is quoted constantly and traces to a Meta deck that exists only as a third-party re-upload — there is no Meta-hosted original anywhere, including archives. The figures Meta does publish on its own site are the 11% and 7.6% above.
The best controlled comparison so far
A quasi-experimental study by researchers at Columbia, Harvard, the Technical University of Munich and Carnegie Mellon, run with Taboola across more than 500 million impressions and 3 million clicks, compared AI-generated ads with human-made ones using a sibling-ads design.
AI ads matched and slightly beat human ads. But the finding worth carrying is the one in the caption — within the AI set, the ads that read as human outperformed the ads that read as generated. The advantage is not in the technology being visible. It is in the technology being cheap enough to iterate with, while the output still passes as made by a person.
What people actually think about it
Gallup surveyed 3,270 US adults in May 2026, with a margin of error of about 2.4 points.
The shape of that is consistent and worth reading carefully. Acceptance falls as AI moves from assisting the work to being the work, and it collapses at synthetic humans. Disclosure helps rather than hurts: the 75% figure is conditional on disclosure.
There is a genuine contradiction in the literature here that I am not going to resolve. One 2025 study of service advertising, in the Journal of Retailing and Consumer Services, finds an AI disclosure lowers trust in the advertiser and attitude toward the ad. Other work finds disclosure raises trust when the AI use is noticed anyway. Both are plausible, and they probably differ on whether the audience would have spotted it regardless. If your AI use is obvious, disclose it. If it is genuinely invisible, the research does not tell you what to do, and anyone who says otherwise is picking a side.
Where this actually breaks
Having shipped a good deal of AI imagery and video into live client campaigns, the failure point is not where people expect.
The stack, since naming it is more useful than not: Synthesia and Arcads for presenter and actor video, ElevenLabs for voice, Google's Nano Banana for image work, and Higgsfield AI and fal for generation across both. It changes as the models change, and it is a working set rather than a recommendation. What does not change is that everything below comes from running these on client campaigns, not from reading about them.
Scripting is solved. It is genuinely good. Editing and orchestration are improving quickly. What still has a long way to go is the footage itself — actual shots, transitions, and holding a shot for a usable length. And the models that produce the best output are the expensive ones, which is where the economics break at scale: the cost per usable second is high enough that volume production stops being obviously cheaper than filming.
The same pattern holds for copy rather than footage. The ten prompts we actually run for Google Ads are written up with the point at which each one stops being trustworthy, because knowing where the output goes wrong is what makes it usable.
The real difficulty, though, is neither generation nor cost. It is building an end-to-end pipeline and iterating on it enough that the output does not read as slop. Anyone can generate a clip. Almost nobody has built the loop that turns clips into something worth running.
For synthetic presenters specifically, the craft fix is to give them less frame. Shrink the avatar and let b-roll and captions carry the composition, or cut the presenter in only between segments of a real product demo. That lines up exactly with the controlled study above: the winning AI ads were the ones that did not look like AI ads.
On the backlash
I think the negative sentiment is accurate, and I would not argue with anyone who holds it. AI slop is real, and some people are more allergic to it than others.
But the worst of it is not produced by the technology. It is produced by people who never built a workflow, never iterated, and spam the output. That will improve — the models are improving, the cost of the good ones is falling, and once strong creators start publishing their workflows the market copies best practice quickly. AI content that is genuinely useful to the person seeing it will be accepted, disclosure and all. Content that exists because it was cheap to make will not, and should not.
One case is worth remembering as a caution about testing. Coca-Cola's AI-generated holiday advertisement scored a perfect 5.9 out of 5.9 in System1's pre-launch testing in both the US and UK — and then met a substantial public backlash on release. Whatever the panel was measuring, it was not the thing that happened next.
What to read next
- Ad creative testing — how to evaluate any of this without fooling yourself.
- User-generated content and creator ads — the format synthetic presenters are usually imitating.
FAQ
How much does generative AI actually lift ad performance?
On Meta's own site, the formal figures for its generative creative tools are an 11% improvement in click-through rate and a 7.6% improvement in conversion rate. Set that beside the adoption numbers — more than a million advertisers, 15 million ads in a single month — and the honest reading is that the technology has transformed how much creative gets made and barely moved how well it works. An 11% click-through improvement is a good result for a production tool. It is not a different category of advertising.
Where does the "32% efficiency from creative diversification" figure come from?
Nowhere citable. It is quoted constantly and traces to a Meta deck that exists only as a third-party re-upload, with no Meta-hosted original anywhere, including archives. The figures Meta does publish on its own site are the 11% and 7.6% above.
Do AI-generated ads beat human-made ones?
Marginally, in the best controlled comparison so far. Researchers at Columbia, Harvard, the Technical University of Munich and Carnegie Mellon ran a sibling-ads quasi-experiment with Taboola across more than 500 million impressions and 3 million clicks, and AI-generated ads reached 0.76% click-through against 0.65% for human-made. The more useful finding is not in that comparison: within the AI set, the best performers were the ads that did not look AI-generated. The advantage is not the technology being visible, it is the technology being cheap enough to iterate with while the output still passes as made by a person.
Should I disclose that an ad used AI?
If the AI use is obvious, yes. Gallup surveyed 3,270 US adults in May 2026, with a margin of error of about 2.4 points: 75% found using AI to brainstorm acceptable when disclosed, 53% accepted AI creating the final content, and 62% found AI-generated people or voices unacceptable. Acceptance falls as AI moves from assisting the work to being the work, and collapses at synthetic humans. Beyond that the literature genuinely contradicts itself — one 2025 study finds disclosure lowers attitude toward the ad by reducing brand trust, while other work finds it raises trust when the AI use would be noticed anyway. If your AI use is genuinely invisible, the research does not tell you what to do.
Where does AI creative actually break in practice?
Not where people expect. Scripting is solved and genuinely good, and editing and orchestration are improving quickly. What still has a long way to go is the footage itself — actual shots, transitions, and holding a shot for a usable length — and the models producing the best output are the expensive ones, so the cost per usable second stops volume production being obviously cheaper than filming. The real difficulty is neither generation nor cost: it is building an end-to-end pipeline and iterating on it enough that the output does not read as slop. Anyone can generate a clip. Almost nobody has built the loop that turns clips into something worth running.
Sources on this page
- Meta, "Meta's AI Product News", September 2024, and Meta's "Demystifying Creative Diversification" page for the 11% click-through and 7.6% conversion-rate figures. Company-reported.
- Google, "Digital Advertising and Commerce in 2026", company blog. Roughly 70 million assets generated in the fourth quarter of 2025. Company-reported.
- Klarna press release, 2024. Self-reported savings and cycle times.
- Columbia, Harvard, Technical University of Munich and Carnegie Mellon with Taboola, 2026, via Taboola's release. 500 million-plus impressions, 3 million clicks, sibling-ads quasi-experiment.
- Gallup for Bentley University, May 2026. n=3,270 US adults, 4–11 May 2026, ±2.4 percentage points.
- Grigsby, Michelsen and Zamudio, "Service ads in the era of generative AI: Disclosures, trust, and intangibility", Journal of Retailing and Consumer Services, 2025.
- Adweek reporting System1 panel data on the Coca-Cola AI holiday advertisement, 2024. Panel score is System1's; the public reaction is qualitative.
- The tool stack named above is my own working set as of August 2026. It is not a ranking, a recommendation, or a paid placement, and no vendor on that list was consulted about this page.
