Do AI video ads perform as well as human-made ads?

What do the main studies actually show?

The most controlled test we found is the Ipsos and Syracuse University study. Researchers took 10 original 30-second ads, reverse-engineered a brief for each, had an AI model write a shot list, and fed it into Sora 2. The AI ads were used exactly as generated, with no human post-production, though some needed several attempts to reach a testable standard¹. Viewers rated both sets as broadly comparable, yet the human-made ads scored higher on Ipsos's effectiveness indices: 14% on the short-term measure and 17% on the long-term measure¹.

Kantar reached a different headline. Its earlier analysis of 356 AI-generated ads found the same spread of results as ads made by traditional means, and Kantar's commentary said AI that is obvious or poorly integrated can hurt performance². That is not a contradiction. Both say the tool does not decide the outcome; the idea and the finish do.

Taboola's sibling-ads design compared AI and human ads from the same advertiser on the same day. Raw click-through was 0.76% for AI against 0.65% for human, and the results were broadly similar after statistical controls³. Taboola is an ad-tech vendor that announced the study, so read it as a platform's finding about clicks, not a neutral audit of sales.

A smaller System1 sample of 18 AI-assisted ads averaged 3.4 stars against 2.3 across its database of 123,000 video ads. The author calls it a small study that is biased toward ads that had already caught the industry's attention⁴.

Why do some AI ads fail?

Ipsos found the strongest AI ads were product-driven and direct, built on formats the brand had used before. The weakest were the ones whose brief asked for emotion, humour or a point of view¹. It also found AI used storytelling in 30% of its ads, against 49% in a wider Ipsos review of ads¹.

Tone and finish matter as much as the model. McDonald's Netherlands removed a 45-second AI Christmas film after a social media backlash, with critics calling it unsettling and cynical. The production company's chief executive said the film took ten people five weeks, a claim NBC reported but did not verify⁵. A cheap model does not mean a cheap or easy ad.

The common thread is the brief. Ipsos says the strongest work, human or AI, came from a solid brief rooted in a human truth¹.

Will viewers notice, and does an AI label cost clicks?

Most did not notice in the Ipsos test: only 25% of viewers of AI ads said they were at least somewhat confident the ad was AI. Even so, 79% said companies using AI should have to disclose it¹.

One field study by researchers at NYU and Emory on the Google Display Network found that labelling an ad as AI-generated cut click-through by about 31.5%, while unlabelled fully AI-generated ads had a 19% higher click-through rate than the human control. The coverage does not state a sample size⁶. Treat that as a signal to test labels, not as a reason to hide AI use, because rules in some markets require a label (India).

Attitudes lean the same way. A Canva survey of 3,547 consumers in seven countries found 78% would rather see ads made by people even if AI could produce better ones. Canva sells design software, so treat it as a vendor survey⁷.

Is there evidence for Indian brands?

The Ipsos panel was American. We did not find an independent Indian head-to-head test of AI and human video ads. Meta's June 2025 launch coverage quotes two Indian advertisers: Avimee Herbal reported a 62% drop in cost per sale from AI-generated backgrounds, and Big Basket reported a 2.7% higher click-through rate on AI-powered catalogue ads. Both are brand-reported results for Meta's own tools, not video-ad comparisons⁸.

How do I test AI ads against human ones?

Hold everything constant except who or what made the ad. Ipsos did this by giving both routes the same strategy and brief¹. A Meta practitioner also warns that more ads are not automatically better: new variants should be distinctly different, and he suggests starting with a few meaningful variations and expanding on results⁹.

• Pick one product, one audience, one offer and one landing page, and give each version the same budget.
• Choose the decision metric before launch: cost per sale or cost per lead, not views.
• Write the stop rule down first, so a lucky first day does not decide the test.
• Use AI where the study says it is strongest, such as direct product work and variations, and keep people at the centre of emotional brand storytelling¹.
• Plan for a label from the start, and check the rules for each place the ad will run.