When AI Advertising Campaigns Work — and Where They Backfire

The useful question for an artificial intelligence advertising campaign is not whether AI can make an ad. It can. The harder question is which part of the media plan should let AI touch the creative at all. A retargeting unit that needs twenty product angles, a personalized ecommerce video, and a holiday brand film are not doing the same job. Judging them by one standard is how teams end up either underusing a good optimization tool or defending a brand decision that never should have left the review room.

The split is clearest in video. Personalized AI video can improve click-through when the viewer is evaluating product relevance or offer fit. Narrative AI video can also trigger rejection when the viewer is evaluating sincerity, memory, childhood, holidays, or a brand’s emotional right to speak. Those are both video ads, but they are not the same advertising problem.

That distinction matters because the performance evidence is real. It is also bounded. AI-generated creative is most useful when the campaign has a clear response goal, enough variation to optimize, and a format where the audience rewards relevance more than emotional authorship.

Where AI Creative Has Earned Its Seat

The strongest case for AI in advertising starts with jobs paid media teams already struggle to do manually: produce more variants, personalize faster, and let the system learn which combinations earn response. In that environment, the audience is not usually asking whether the creative feels handcrafted. They are deciding whether the ad is relevant enough to stop, click, compare, or buy.

A Marketing Science study from MIT IDE involving 21,000 consumers found that AI-generated personalized video ads produced a 9.4% higher click-through rate than personalized image ads and a 6.5% higher click-through rate than generic video ads. [1] That is the kind of result a media team can use, especially when video production is the bottleneck and personalization is otherwise limited to a few audience-level swaps.

The result should not be treated as proof that every AI video will beat every human-made video. The study was a single-exposure experiment, so novelty may have contributed to the lift. As AI video becomes more common, some of the early attention advantage may fade. Still, the mechanism is credible: the creative did not win because it was AI; it won because AI made more personalized video variation feasible.

The same pattern shows up in ad copy, only with a smaller effect size. A Columbia Business School, Harvard, Technical University of Munich, and Carnegie Mellon field experiment found that AI-created ad copy generated a 0.76% CTR compared with 0.65% for human-written copy, a statistically significant but modest lift. [2] That is useful. It is not a creative revolution by itself. A paid media lead would read that as incremental advantage in a testing system, not as permission to fire the copy desk.

The important part is not the romance of machine-written copy. It is throughput. If a campaign needs ten headline angles, five benefit frames, three urgency levels, and landing-page alignment across multiple audiences, AI can help teams get to a larger testable set faster. The gain compounds only if someone is still pruning weak claims, checking tone, and connecting the copy to the actual offer.

Dynamic creative optimization makes the same argument in a more programmatic setting. StackAdapt reported that AI-driven DCO delivered 32% higher CTR and 56% lower CPC. [3] Those numbers are attractive, but they come from a platform with a commercial interest in AI advertising adoption, so they are better read as directional campaign evidence than as an independent benchmark every advertiser should expect to reproduce.

Even with that caveat, DCO is one of the cleaner fits for AI because the task is structurally aligned with the tool. The system can rotate product imagery, offers, calls to action, audience signals, and message variants around measurable outcomes. The creative surface is usually functional. The viewer is not being asked to believe that the brand understands their childhood. They are being asked to notice a relevant product or offer.

The Performance Lift Is Real, But It Is Not Magic

Performance marketers do not need AI creative to be magical. They need it to reduce production drag without making the account harder to manage. A small CTR lift can matter when it applies across enough spend, enough placements, and enough iterations. A lower CPC can matter when the conversion rate holds. The trap is pretending that a metric lift in one environment carries the same meaning in another.

A 0.76% CTR beating 0.65% is a measurable advantage, but it is not an argument that AI copy has better judgment than a strong creative team. [2] It says that in a specific experimental setting, AI-written copy produced more clicks. That could come from clearer phrasing, better alignment with platform conventions, greater variation, or the simple fact that many human-written ads in live accounts are rushed, recycled, or overconstrained.

The better operating lesson is to use AI where the feedback loop is fast and the cost of being slightly wrong is low. A weak headline can be paused. A bland product video can lose an auction and disappear. A display unit with the wrong benefit hierarchy can be replaced after the first readout. These are not consequence-free decisions, but they are recoverable inside normal campaign management.

That is also why the most practical AI workflows do not remove people from the loop. They move human effort upstream and downstream: define the creative territory, generate many safe variants, then review performance and brand implications before scaling. For teams building this kind of workflow, a governed process matters more than a single generation trick; an AI creative advertising playbook is useful precisely because the work crosses media, creative, legal, and brand review.

Why Brand Campaigns Backfire Differently

The backfire cases are not just examples of people disliking new technology. They show what happens when a brand asks synthetic creative to carry emotional meaning that audiences expect to be earned.

Coca-Cola’s 2024 AI remake of “Holidays Are Coming” drew substantial negative consumer sentiment, with coverage describing the work as being perceived by many viewers as a cost-cutting move rather than authentic creative expression. [4] The problem was not simply that AI appeared on screen. The campaign touched a piece of brand memory. In that context, visible artificiality changes the read of the entire decision: the audience sees not efficiency, but a shortcut taken through something they thought the brand was supposed to protect.

The Toys R Us “The First Toy Made with Sora” film landed in similar territory. Carma recorded that brand-positive sentiment dropped sharply after the film’s release. [5] The film asked viewers to accept AI-generated nostalgia around a childhood retail brand. That is a high bar. If the audience reads the execution as uncanny, hollow, or opportunistic, the sentiment drop is not a misunderstanding. It is the campaign result.

This is where video becomes the dividing line rather than the answer. The MIT result suggests personalized AI video can lift CTR when the video is doing a functional job. [1] The Coca-Cola and Toys R Us cases show that narrative AI video can become a liability when it is asked to perform sincerity. The format is not the deciding factor. The audience’s evaluation mode is.

Paid media teams feel this distinction in review meetings. Nobody spends three weeks debating whether a retargeting thumbnail felt emotionally earned. But a holiday film, founder story, inclusion campaign, or childhood-memory spot invites a different kind of scrutiny. The creative is no longer only an acquisition asset. It becomes evidence of what the brand values.

Perceived Artificiality Is the Real Drag

One of the more useful findings in the research is also one of the most uncomfortable for teams hoping this is just a labeling problem. The Columbia and Harvard research found that ads that merely appear AI-generated perform worse regardless of whether AI actually made them. [2] In other words, authorship matters less than perception. If the ad looks synthetic, the audience can penalize it even when a human team made the work.

That has two consequences. First, “we used AI” is not the only risk. Bad compositing, uncanny faces, over-smoothed product shots, generic emotional scenes, and overly polished but weightless copy can all trigger the same reaction. Second, “a human approved it” does not solve the problem if the work still reads as synthetic in a context where the audience expected care.

The consumer trend line makes that harder to ignore. EMARKETER reported that 39% of Gen Z consumers actively dislike AI-generated ad creative. [6] Consumer comfort with AI in advertising also fell from 60% in 2023 to 46% in 2024, a 14-point decline. [6] Those are attitude measures, not proof that every AI ad will underperform behaviorally. But attitudes shape the margin of error, especially for brand campaigns where interpretation is part of the product.

This is also why teams should separate “AI-generated” from “AI-looking.” An AI-assisted ad can perform well if it looks useful, specific, and credible. A fully human-made ad can suffer if it looks like a generic synthetic asset. The practical review question is not only what tool made the asset. It is what the audience will think happened, and whether that thought helps or hurts the job of the ad.

Disclosure Solves One Problem and Creates Another

Disclosure is not a clean escape hatch. BCG X reported that 69% of consumers feel manipulated when brands use AI for advertising without disclosure, citing Smartly.io research. [7] The same source reported that 58% of U.S. advertisers now use disclosure labels. [7] The direction is clear even though the 69% figure comes from a commercially interested ecosystem and the survey methodology details were not independently verified in the brief.

The tension is that disclosure can protect trust while also heightening the perception of artificiality. If a viewer already sees the ad as functional — a product comparison, a dynamic offer, a localized sale message — a label may be relatively easy to absorb. If the ad is trying to make the viewer feel childhood wonder or holiday warmth, the same label may push the audience to inspect the work for shortcuts.

That does not mean brands should hide AI use. It means disclosure cannot carry the whole trust burden. The creative still has to be appropriate to the emotional stakes of the campaign. For more on the broader consumer trust problem around AI marketing, the related analysis on AI-generated marketing and the trust gap is a useful companion to the paid media evidence.

A Decision Standard for Paid Media Teams

The cleanest operating rule is to start with what the audience is evaluating. If the viewer is evaluating relevance, speed, offer fit, product utility, or comparison value, AI creative can be a strong production and optimization layer. If the viewer is evaluating sincerity, memory, belonging, identity, grief, celebration, or the brand’s emotional right to speak, AI should stay subordinate to human creative judgment.

• Use AI aggressively for variant generation when the campaign has a measurable response goal and rapid feedback loop.
• Use AI carefully for personalized video when the personalization improves product relevance rather than pretending to create intimacy.
• Treat vendor performance claims as directional unless the measurement method, baseline, and campaign context are clear.
• Review AI-assisted work for perceived artificiality, not only for whether AI was technically used.
• Keep humans in charge of the core idea for nostalgia, holiday, founder, purpose, and identity campaigns.

This standard is stricter than “AI for performance, humans for brand,” because plenty of brand-safe performance work still needs taste and plenty of brand campaigns can use AI somewhere in the process. AI can help with animatics, resizing, localization, storyboards, internal concept exploration, or production cleanup. The boundary is the part of the work the audience is being asked to believe.

A synthetic product demo that helps someone understand fit, price, or use case can earn its place in the media plan. A synthetic holiday memory that asks the audience to feel warmth toward a brand is taking on a different burden. The first is judged by usefulness. The second is judged by sincerity. An artificial intelligence advertising campaign can succeed in both production environments, but it should not be given both emotional jobs.

References

1. Frontiers: Generative AI and Personalized Video Advertisements, MIT IDE / Marketing Science
2. AI-created ad copy field experiment, Columbia Business School / Harvard / Technical University of Munich / Carnegie Mellon
3. AI in advertising: How to use it the right way in 2026, StackAdapt
4. Coca-Cola “Holidays Are Coming” AI remake backlash coverage, multiple news outlets, 2024
5. Toys R Us “The First Toy Made with Sora” brand sentiment tracking, Carma, 2024
6. 2026 survey on consumer attitudes toward AI-generated ad creative, EMARKETER, 2026
7. How AI Is Reshaping Advertising for the First Time in a Decade, BCG X, 2026