How generative AI can shred a brand

The AI isn't shredding your client's brand. The absence of a rule is. Brand guidelines held fonts, colour, spacing and message together for decades, and nobody has written the equivalent for generation — so every designer prompts differently, and a brand that took a decade to build starts drifting inside a quarter.

Forty years on the creative side. A few agencies of my own, creative director at others, twenty years consulting brands on strategy, hundreds of high-end photo shoots.

On a national campaign you might shoot thirty angles. The work isn't getting a good frame — you'll get plenty. The work is the days spent making all thirty look like one campaign. Same light, same grade, same distance from the subject, same restraint about what's in the background. That's what a brand looks like from the inside: not one great image, but thirty images that agree with each other.

That agreement is what's going.

Open any real brand book and it reads like a specification, because that's what it is. The typeface, licensed. The hex values, exact. The clear space around the mark, measured, so nothing crowds it. What the logo may sit on and what it may not. How the voice sounds in a headline versus a footer.

None of that was decoration. All of it existed so that a designer in Manchester and a designer in Dallas, working three months apart and never speaking, produced work that looked like it came from the same company.

And imagery had its own chain. A shot was licensed from a stock house or commissioned on a set, the licence had terms and territories, and the whole thing went up the chain and back down — art direction, client, legal — before it went anywhere near a media buy. That chain was slow, everybody complained about it, and it was doing something. Every step was another chance to catch a piece of work that didn't look like the brand.

Now the chain is a prompt and a campaign upload. Go and open your client's brand book, and count how many of its rules a generated image can actually be held to.

Two designers, same brief, same model, same afternoon — and two different brands come out. Not because either is careless. Because a prompt is a description, and a description is not a spec.

Image-conditioned generation — feeding the model an actual reference — holds closer to the reference than text prompting does, and LoRA fine-tuning captured the high-level look but lacked precise consistency. That was measured on architectural images rather than branded campaigns, so take it as a direction rather than a rate: give the model an actual reference and it holds, describe the reference in words and it wanders.

Now hold that against the three of those a generated image has to carry on its own.

Colour. I couldn't find a controlled test showing that a current generator reliably lands a specified hex across repeated runs. Putting the value in the prompt isn't colour management — you're getting a pixel that's been through lighting, gradient and compression on the way out.

Type. Getting legible text out of an image model is still an open research problem, and legible is a long way short of the licensed typeface your client paid for, set at the right weight with the right kerning.

Consistent imagery across a set. Holding many references together is unsolved enough that researchers built a benchmark for it this year. Nobody benchmarks a solved problem.

Colour, type, and thirty images that agree with each other. Three of the things a brand book exists to fix, and generation can promise you none of them.

I went looking for the evaluation — something that takes a real brand system, generates a realistic campaign set, and scores it on colour, logo, type, layout, composition and tone. I couldn't find one.

Not from the tool vendors, not from an independent lab. There are vendor claims about how many reference images it takes to teach a model a style. None of them is an independent measurement, and they don't agree with each other.

So when a platform tells you it keeps your client on brand, ask what that was measured against.

Nobody set out to loosen their brand. What happened is that generation got fast and cheap, and fast and cheap are extremely persuasive to a marketing director under pressure to ship more.

Every client I talk to wants more assets out the door this quarter than last. That isn't automatically good. Volume is not the same as presence, and a hundred pieces that half-look like the brand do more damage than ten that are exactly right — because the hundred are what the market now thinks the brand looks like.

Every loose system I've walked into went first. A tight book degrades slowly because there's something to measure drift against. A loose one has nothing to hold the line, and generation finds the gap immediately.

As a company grows, the brand stops being a look and becomes something people are paying for — the reason a customer picks you at the same price, and a line an acquirer is partly buying. Letting it blur to save a week is an expensive trade nobody writes down.

The fix isn't to ban the tools. The production savings are real. The fix is that generation needs the same kind of document brands always had — specific, enforceable, and written down where a freelancer can read it.

What it should settle:

Where generation is allowed, and where it isn't. Concepting, variants, backgrounds, mood — probably yes. The hero image of a national campaign, the founder's portrait, anything a lawyer signs — probably not. Name the surfaces, don't leave it to judgment.

What every prompt must carry. If two people write it differently you'll get two brands, so fix the structure — the reference images, the required and forbidden attributes, the treatment of light and camera distance, the things that never appear in frame. A prompt template is the new type spec.

What must be applied after generation, not during it. The mark, the exact hex, the licensed typeface. Those go on in your design software, where they come out the same every time — because that's the only place they currently do.

Who approves, and against what. Not whether it looks nice — whether it sits beside last quarter's work without an obvious seam. Two people, same rubric.

What gets recorded. Which model, which references, which prompt. If you can't reproduce it in six months, you don't have an asset, you have a lucky result.

Think of the campaigns you actually remember. The Macintosh launch. Whatever your equivalent is — the one that made a category feel different afterwards.

None of that was won on execution. It was somebody deciding what the thing meant and what to point the camera at. The craft downstream mattered enormously, but the decision came first, and no prompt produces it — because a prompt can only describe something that already exists in somebody's head.

The idea is still yours. What's changed is that the machinery downstream of it will now drift off it while you're not looking, faster than any junior ever did.

Take the last campaign you shipped with generated imagery in it. Pull ten of the images into one document, at the same size, side by side.

Then look at the light. The colour temperature. How far the camera sits from the subject. How busy the backgrounds are.

If they don't look like they came from the same shoot, they don't look like they came from the same brand — and that's what your client's customer is seeing, in a feed, next to a competitor who still runs a photo shoot.

Twenty minutes. Then go and write the guideline, because the exercise only tells you it happened — the guideline is what stops the next one.

Fair disclosure, I sell a platform that does part of this, so I'm biased about it. Good luck with the audit — I hope it's duller than mine was.

Can generative AI stay on brand?

Partly, and not on the parts a brand book cares most about. A peer-reviewed 2026 Springer study comparing image-control workflows for architectural design found that image-conditioned generation holds closer to a supplied reference than text prompting does, while LoRA fine-tuning captured the high-level look but lacked precise consistency. What I could not find published anywhere is an evaluation showing a current tool holding an exact specified colour across repeated runs, or setting a licensed typeface natively. Those get applied after generation, in your design software, where they come out the same every time.

Why do two designers using the same AI tool produce different-looking work?

Because a prompt is a description, and a description is not a specification. A brand guideline fixes typefaces, hex values and clear space so two people who never speak produce compatible work; a prompt fixes none of that unless somebody writes the structure down. A 2026 Springer study comparing image-control workflows found text-only prompting drifted further from a supplied reference than image-conditioned generation did, and that LoRA fine-tuning captured a high-level look without precise consistency.

Has anyone measured how well AI tools follow a brand system?

I went looking and couldn't find one — no evaluation that takes a real brand system, generates a realistic campaign set, and scores it across colour, logo, type, layout, composition and tone. Holding multiple references together is unsolved enough that researchers introduced a benchmark for it in June 2026, OmniRef-Bench. Ask any platform claiming brand adherence what it was measured against.

• Five places I still won't let AI near client work — the copyright and disclosure exposure underneath all of this.
• The ad you approved isn't the ad that ran — what Meta and Google change after you've signed everything off.
• GEO vs SEO: what the hell is the difference? — what the engines read about a brand before they answer.