How to use AI in your creative without quietly bankrupting the one thing you can't buy back

I've been staring at AI-generated product images for what is frankly an unhealthy amount of time this week, and I've arrived at a conclusion that I think the industry needs to hear: we have collectively lost the plot on where the line is.

Not the "is AI good or bad" line. That argument is over. Almost every marketing leader is now using AI somewhere in their daily creative work, and nearly all of them plan to spend more on it. The tools are in the building. The interesting question stopped being whether and became where—which parts of your creative can safely pass through a model, and which parts you're touching at your own risk.

Here's what makes this genuinely spicy: while adoption hit near-saturation among marketers, roughly half of US consumers say AI has made the content they see worse. That figure climbs higher among younger buyers. So the people making the content have decided AI is brilliant, and a large chunk of the people receiving it have decided the opposite. The gap between those two positions is where brands go to die. (Quietly, expensively, and usually without realising until the metrics start looking weird.)

I've read approximately nine thousand frameworks for "responsible AI in creative" and most of them are elaborate ways of saying nothing. So here's the one that actually works, and it fits in a sentence:

Is AI changing how something real is presented, or is it manufacturing the evidence your customer uses to decide whether to trust you?

That's it. That's the whole thing.

Changing the lighting on a real product photo? Treatment. Generating a fake human to look into the camera and tell someone your serum changed their life? Manufacturing evidence. The closer you drift toward manufacturing evidence, the faster you spend down what I'd call your brand's trust reserve—the accumulated sense that you're a real company selling a real thing.

That reserve refills slowly and drains fast, and a single obvious misfire can cost more of it than months of careful work put in. Which brings us to the part where I tell you what's safe and what isn't. (Don't worry, there's good news first.)

Remixing assets you already own. Hilton used image-to-video AI techniques on existing property photography across roughly 700 locations rather than reshooting each one, reporting around 42% more reach, an 8% lift in ROAS, and about 13% more bookings. Nothing material about the product changed. Only the treatment did. The customer is still looking at the real hotel—just presented more dynamically than a static photo ever could.

Speeding up the mechanical layer. Unilever's AI Studio reportedly develops creative assets about 30% faster, with its oral-care brand Closeup producing over 100 assets across 10 setups in three days. The human decides what the campaign is, the model absorbs the repetitive work of rendering it in twenty sizes and five formats. Strategic decisions stay human. The labour doesn't.

The deliberately impossible. Koia used clearly synthetic, animated-style creative for its Costco protein-powder launch on Instagram Reels and hit around 48,000 organic views—roughly 6.4x its usual reach. It worked precisely because it was obviously unreal. Nobody accuses Pixar of fraud. Deception requires a claim to reality. Openly impossible imagery makes no such claim, which is exactly why it's safe. (Also, it tends to perform well because it's actually interesting to look at, which is more than you can say for most product photography.)

Ideation. The safest of all, because the customer never sees it. Using a model to generate directions and pressure-test concepts before committing production money is close to pure upside. It's scaffolding, and scaffolding comes down before anyone moves in.

Synthetic humans in trust-dependent contexts. There is a meaningful difference between using AI to build a product photography template and using an AI-generated person to look into the camera and recommend your product. Goodwipes drew the line exactly here: comfortable with AI-optimised static visuals from existing shoots (reporting above-1% CTR on those assets), but unwilling to put an AI-generated spokesperson on screen for a product tied to a private, trust-sensitive routine. If your category asks the customer to trust you with something personal, a manufactured human is the fastest way to spend down that trust.

The perception penalty. REI got publicly roasted after Meta's Advantage+ AI tool auto-altered a Van Rysel bike ad to include a bicycle with two sets of handlebars, misrouted chains, and gibberish frame text—the kind of tell that turns a campaign into a meme. (The ad ran for a week before anyone at REI noticed. Reddit, naturally, noticed immediately.) But here's what's more instructive: Quip was accused of using AI on a stylised campaign when they hadn't, simply because it looked like it might have. The market has started pricing artificiality directly, through engagement drops and comment sections that turn forensic. An advantage that works only while the customer doesn't notice is a bet that recognition never arrives—and recognition is getting easier, not harder.

The sameness trap. When every brand reaches for the same tools, the tools reach back with the same fonts, the same aesthetics, the same vaguely aspirational lifestyle imagery. As AI-mediated discovery grows, the undifferentiated brand is the one most easily substituted for a cheaper equivalent. The cost doesn't show up this week. It shows up as the slow erosion of any reason to prefer you. (Which is, coincidentally, the most expensive kind of cost there is.)

Letting the tool set the direction. A model can generate a campaign, but it cannot decide what your brand means. If you let it try, it will hand you back an averaged version of your category rather than a sharpened version of you. The discipline is keeping AI on the execution side of the wall and your brand's voice, story, and judgement on the human side. The tool should not be dictating your point of view back to you.

Here's where things get properly interesting—and by interesting I mean "the kind of interesting that requires you to sit down."

As of August, the EU's transparency rules and California's AI transparency law both took effect on the same day, requiring large AI providers to embed machine-readable provenance into the images, video, and audio they produce. Leading image generators now embed invisible watermarks paired with content credentials that record which tool produced a file and when. Claude now weaves a statistical watermark into the copy it generates—applied worldwide with no opt-out, surviving copying, pasting, and light editing.

The practical consequence: the creative you generate with these tools increasingly arrives pre-marked, carrying a signal a platform or regulator could in principle read, whether or not you say a word about it. Your creative is carrying a receipt.

This is why the treatment-versus-evidence distinction matters more than any specific disclosure rule. If AI is changing how something real is presented, a provenance mark is harmless because there was never anything to conceal. If AI is manufacturing the evidence a customer relies on to trust you, the mark becomes a liability—and the mark is becoming automatic.

Now, I can already hear the objection forming, and it's a fair one: "That's lovely, Jo, but I don't have a photographer, a studio, or a production budget. AI is the only reason I have creative assets at all."

And you're right. For a seller at the early stage—bootstrapped, running lean, reinvesting every spare pound into inventory—AI-generated creative isn't a compromise. It's a genuine unlock. The alternative isn't a glossy photoshoot. The alternative is no creative at all, or creative so bad it actively hurts conversion. In that context, a well-executed AI image is a massive step up, and nobody should feel guilty about using one.

The nuance is in what happens next. The moment a brand starts generating real revenue, the instinct is to reinvest into scaling—more spend, more SKUs, more markets. Rarely does "invest in real creative assets" make the shortlist, which is a shame, because it might be the single highest-leverage thing you can do as you grow.

Here's why. Outside of pure commodity categories, purchasing decisions are emotional. People buy brands they believe are authentic. Not brands that are authentic in some abstract philosophical sense, but brands that feel real: real people, real stories, real photography of a real product that someone actually touched. That feeling is extraordinarily hard to manufacture synthetically, and it becomes more valuable the more the market fills up with content that obviously came out of a model.

Gozney is the instructive case here. Four-figure pizza ovens, a roughly 30-day consideration window, over $100m in revenue, and a 400,000-subscriber YouTube channel built on real craft and real storytelling. They restrict AI to internal drafting and keep it out of anything a customer sees—because their audience is partly paying for the human hand behind the product. When their elaborate Tonester paint collaboration shoot was so visually striking that viewers assumed it was AI-generated, that told you everything about where perception is heading: visibly real work is becoming a differentiator in its own right.

The trajectory for smaller sellers looks something like this: AI creative gets you started at a level that would have been impossible five years ago without a proper budget. That's genuinely transformative. But as you grow, scaling your ad spend from £5k to £10k might well be the right call—and carving out part of that budget for authentic creative, real influencers, and genuine brand ambassadors is just as important. The brands that build a library of real assets no competitor can replicate with a prompt are the ones that compound. The spreadsheet never tells you to buy authenticity. It's also the highest-returning line item you'll never see in a ROAS column.

The line between safe and unsafe AI creative is not fixed, and it will keep moving. What holds steady is the underlying test: is AI changing how something real is presented, or manufacturing the evidence a customer relies on to trust you? The first is where the efficiency lives. The second is where the brand gets spent.

If you're early-stage and AI creative is the only creative you can afford—use it, stay in the safe zone, and don't apologise for it. But the moment you have margin to reinvest, real creative assets aren't a luxury. They're a growth lever—possibly the most underrated one available, precisely because the spreadsheet never tells you to buy it.

The harder question—and the one worth sitting with over whatever you're drinking right now—is not which tasks you're willing to automate. It's which signals you're willing to let a customer eventually discover were automated. And whether your brand is still standing when they do.

🤔 FrequentlyAskedAI - Instant AI-powered FAQ builder

👨‍💼 Save time by My Ask AI - Less time on customer support, more time on customer success.

📆 Vimcal - Calendar app on steroids. Assists users in scheduling and organizing events efficiently

🤖 Getverk AI - It allows its users to hire AI-generated employees.

😍It was great, all I needed

🙂Good, not great

🙄It sucked