BlogAI and your brand
What AI can actually do for your brand, and what it should never touch
6 min read

REI's approved photo showed a cyclist on a bike with one set of handlebars. In June the version running in its ads had two, and the cyclist hired for the shoot had a different face: Meta's AI tools had changed the image (GadgetReview). Neither change was in the file REI's team approved. Had it happened after 30 October, REI would have owned both: under Meta's revised terms, in force since 30 October 2026, advertisers assume responsibility for AI-generated output (ContentGrip). The UK's advertising regulator put it more bluntly: "delegating production to an algorithm does not delegate accountability" (Lewis Silkin).
That is the trade at the centre of AI advertising this year. AI solved the problem that dominated the last decade, which was making enough ads. It created a new one, which is deciding what stays out of the machine's hands.
The problem AI solved
Winners are rare and short-lived. In Motion's 2026 benchmark of $1.29 billion in Meta spend, fewer than one creative in ten became a winner (Motion), and each one wears out in weeks. Finding the next one means trying different hooks, prices and offers for each audience, country, language and season, which is why, in AppsFlyer's 2025 report, the biggest game advertisers made about 2,700 creatives a quarter (AppsFlyer).
Before generative models, each of those files cost someone's hours. Now they cost minutes: 56% of the top-100 grossing mobile games already used AI to make ad assets in 2025 (Liftoff, citing AppMagic). The volume problem is largely solved. Five others came with the solution, and they decide who wins now that everyone has the same models.
Five problems that came with the volume
Fatigue, at everyone's expense. When every advertiser can make far more ads, people see far more of them. Ad impressions in gaming grew 20% in 2025 while in-app purchase revenue grew 1.3% (Segwise, citing AppsFlyer and Sensor Tower). Game Growth Advisor calls it the AI creative paradox: each studio's extra volume is rational, and together they bid up the same attention, so volume stops being an edge.
Sameness. Everyone uses the same models, so the extra ads look alike. Freddie Campbell of House of Communication told the IPA in September that access to models no longer sets anyone apart, and Sara Chapman of adam&eve\TBWA warned of "creative collapse" as models trained on AI-made work drift toward the average (IPA). In-house teams feel it first: the ANA's 2026 report puts creative stagnation, not workflow, at the top of their threat list.
Distrust. 57% of people think brands use AI to take shortcuts (IPA). What they punish, though, is carelessness rather than AI itself: Kantar's study of 356 ads found no link between AI use and effectiveness, and AI used in ways viewers didn't notice performed better (CO/AI, citing Kantar).
Approved details change on their own. A generative model doesn't edit a shot, it draws the whole thing again, and at this volume nobody watches every frame. REI's handlebars were one case. In October, Bookshop.org found an AI-generated illustration inside an ad its own people had made, after an automatic account update enrolled it in Meta's generative features, and the agency Exverus pulled live ads after finding translated copy and generated images it hadn't chosen (ContentGrip).
What the team learns doesn't stay. In most AI tools each batch starts from a fresh prompt, and in a platform's own automation the learning stays with the platform. Brainlabs' Liz DeAngelis warned in November that brands letting Meta keep adjusting their creative may see results improve "without knowing what actually improved it" (PPC Land); Don't rebuild. Build on takes that argument further.
Where the line goes
The five problems sort into three kinds. Fatigue is a volume problem, and volume is what AI does well. Sameness, distrust and changed details come from decisions nobody made, so those decisions stay with people. Lost learning is a memory problem, so whatever the team decides has to survive from one batch to the next.
There's early evidence for the split. In a preliminary Google Display Network study cited by Liftoff, ads AI generated from briefs written by people beat ads built by human experts alone by 19% in click-through, while AI edits of existing human ads showed no lift (Liftoff).
AI makes the volume. New ads every week, in every size, language and offer, is repetitive work at a scale no team can staff, and its result is easy to check. One approved ad laid out again for every placement, its text translated and reflowed for every market, its price changed in every version, and as many openings as a test needs: that is the work to hand over.
People keep the choices. What the ad is about, who it speaks to and which take goes out. WPP's Daniel Hulme calls this the Monet Problem, since "AI can spin up a thousand renditions of an idea in a minute" and somebody still has to choose well (IPA).
Some things don't change once they're approved. Gameplay shows players the game they'll download, and in November the ASA banned install ads from King and Scopely for showing things players wouldn't find (MobileGamer.biz); that is the subject of why your own footage beats a generated ad. Licensed characters have to appear exactly as the licensor signed them off: even when Disney licensed more than 200 characters to OpenAI's Sora, it kept its talent's voices and likenesses out of the deal (Axios). A disclaimer, an age limit or a rate has to run as legal approved it, and since 2 August 2026 Article 50 of the EU AI Act has required deployers to disclose deepfakes, meaning realistic AI-generated or manipulated images, audio or video of real people, objects, places or events, with a lighter duty for evidently artistic work and fines of up to €15 million or 3% of worldwide turnover (The Drum). And the price is a business decision: in November 2025 Meta's AI promo feature switched on a 100%-off offer by itself (AdExchanger).
Where the line sits in practice
Drawing the line is easier when the ad itself is built in two layers. In Motn, footage, gameplay and licensed art go in and stay as they are, frame for frame. The headline, price, offer, logo, end card and legal lines sit on top as live text and motion graphics, written as code by Claude, so a new size, language or offer changes only that layer, and a fix changes one word without moving anything else. Version 20 is the approved footage plus changes anyone can read, so checking it takes minutes rather than a frame-by-frame hunt.
Nothing is in the exported file unless the team put it there. Meta's enhancement settings still need checking at publishing, because they act after upload, whatever tool made the file.
The third part, memory, is what turns the line into a habit. Say a puzzle game's team chose the opening, the board shot, the end card and the legal line for its Halloween event. For Thanksgiving it swaps the event art and the offer, and everything else it decided in October comes along; the opening that won leads the next test, and the losers stay on record so nobody pays to test them again. How that works week to week is in the weekly refresh.
To see the two layers on a real ad, book a demo, or see Generated, then editable at every step on the homepage.


