Do AI ads actually work?

The strongest study yet on this question found that human-made ads outperformed AI-made versions of the same brief by fourteen per cent on short-term sales impact and seventeen per cent on long-term brand equity. The same study found most viewers could not tell which was which. Both halves matter. Consumers did not detect the AI and bought less anyway.

Sales impact

+14%

Brand equity

+17%

Detection

25%

We make AI films for a living. This is the research that argues against us, and it is better research than anything currently arguing for us.

What the test did

Ipsos, with the Newhouse School at Syracuse University, published the work on 18 May 2026. Twenty ads across ten brands: Cheerios, Chewy, Febreze, Fiat, H&M, Old Navy, Herbal Essences, Ray-Ban Meta, TurboTax and Visa. Each pre-2021 human-made thirty-second spot was paired with a fully AI-generated counterpart. Gemini reverse-engineered the brief and the shot list, Sora 2 generated the video, and there was no human creative input on the AI side. Three thousand US consumers, tested in Ipsos Creative|Spark.

Human ads averaged fourteen per cent higher on the Creative Effect Index, Ipsos's measure of short-term sales impact, and seventeen per cent higher on the Equity Effect Index, its measure of long-term brand health. On detection, only twenty-five per cent of the people who watched an AI ad were even somewhat confident it was AI, and forty per cent of all viewers were unsure either way.

Asked to compare directly, people rated the AI ads less creative, less emotionally engaging and less informative. They rated the human ads more entertaining, more unique and more talkable. Adam Peruta, a co-author, framed it in Harvard Business Review as ads that perform worse even when customers cannot tell them apart.

Where the deficit actually sits

One number in the study explains most of the others. Left to construct an approach on its own, the AI pipeline chose storytelling only thirty per cent of the time, against forty-nine per cent in Ipsos's separate norms for human advertising. It did not fail at execution. It declined to tell a story.

That shows up in which brands it handled well. AI did best on product-driven, direct briefs, Febreze and Herbal Essences, where the job is demonstration. It did worst where the job was storytelling, emotion and point of view. The deficit is not in the pixels. It is in the decision about what the film is.

The caveats, stated rather than buried

Ipsos ran the study on its own commercial testing products, so treat it as semi-independent rather than vendor-neutral, with the mitigation that Ipsos has no stake in AI production either way. It tested one pipeline, Gemini plus Sora 2, with no human creative input, which is not how anyone ships paid work. And the two releases disagree with each other: the Syracuse announcement says thirteen per cent of viewers were confident and reports deltas of plus eleven and minus five, while the Ipsos report says twenty-five per cent and plus fourteen and plus seventeen. The figures quoted above are the Ipsos ones, named to their metric.

The counter-evidence

System1 reached close to the opposite conclusion. Reported in Marketing Week, it tested eighteen AI-assisted video ads against its database of a hundred and twenty-three thousand. The AI-assisted ads averaged 3.4 stars against 2.3 for the database, and UK AI ads scored 4.6 against a 2.6 UK average. Coca-Cola's 2025 "Holidays Are Coming" hit System1's maximum 5.9 stars, with seventy-one per cent of emotional response registering as happiness against a thirty-six per cent category average. Ads that viewers recognised as AI scored better than ads they did not, 3.3 for the top half against 2.3 for the bottom.

“The machine is good at what was always mechanical and weak at what was always authorship.”

Four problems with it, and we would raise all four if a client cited it at us. Eighteen is a very small sample. The selection is biased toward ads that had already attracted industry attention, so it is closer to a sample of famous AI ads than of AI ads. Seven of the eighteen were made by Jellyfish and Brandtech Group companies, the same groups whose tooling was doing the LLM scoring. And the human-to-LLM correlation was 0.56, which is not a strong enough agreement to carry a star rating. We were unable to establish a publication date for the study.

The middle finding, which is probably the real one

Kantar analysed three hundred and fifty-six ads across TV, digital and social and found no correlation at all between AI usage and creative effectiveness. AI-involved ads turned up among both the most and the least effective for brand recall. Kantar sells creative testing, so the same caution applies as to Ipsos and System1.

Its facial-coding analysis from 4 November 2025 is more specific and more useful. Generative AI ads evoked stronger emotional reactions but lower net positivity and lower branding on average, and the damage was concentrated where the AI use was obvious. Where AI was used seamlessly, over forty per cent of ads landed in the top tier for branded cut-through. Kantar also found that forty-one per cent of consumers say AI-generated ads bother them, against twenty-nine per cent of marketers.

What consumers say when you ask enough of them

The largest sample we found is DoubleVerify's, published 29 July 2026: twenty-two thousand consumers across twenty-two markets, plus two thousand and twenty marketers across twenty-one. Forty-two per cent said a brand's use of low-quality or uncanny AI advertising would negatively affect their opinion of that brand. Forty per cent view polished, professional AI advertising positively. Fifty-six per cent say they cannot consistently identify AI content. Regional split worth noting for anyone working in this part of the world: fifty-three per cent in APAC were positive about advertising inside AI chat platforms, against thirty-four per cent in North America. DoubleVerify sells AI-related verification.

Running underneath all of it is a gap in who believes what. The IAB framework reports that eighty-two per cent of advertisers think consumers feel positive about AI-generated advertising, against forty-five per cent of Gen Z and Millennial consumers who actually do. Thirty-seven points, and it has widened since 2024. Pew has seventy-six per cent of US adults saying it is extremely or very important to be able to tell AI from human content. Among Gen Z, thirty-four per cent read an AI-using brand as creative and thirty per cent read it as inauthentic.

What to do with this on a brief

The evidence does not say AI advertising fails. It also does not say it is cheap. What it says, consistently across three testing companies with different incentives, is that the measured deficit sits in storytelling and brand equity rather than in craft detection. Audiences mostly cannot spot the method and still respond less to the result.

That is a description of a job, not a verdict on a tool. The machine is good at what was always mechanical and weak at what was always authorship. So the practical read is narrow and testable: product-driven, direct, demonstration-led briefs are where the published evidence supports AI production today. Anything whose job is to make somebody feel something about a brand in thirty seconds should be written, cast and directed by people, whatever generates the frames. If a studio tells you otherwise, ask them which of these four studies they are working from.