The current pushback against AI-assisted branding is being described as a rejection of AI.
I think that is the wrong read.
It is mostly a rejection of weak constraints.
Why this is a backlash
June’s branding coverage is useful because it shows the mood turning. Creative Bloq’s feature on everything looking the same and Design Week’s piece on blanding both argue that AI is accelerating a pre-existing sameness problem rather than inventing a new one.
Creative Bloq’s AI rebrand coverage makes the other half of the point: cautious adoption beats blind enthusiasm.
That is exactly right for branding work.
What goes wrong in AI rebrands
The weak version of AI-assisted branding asks the system to generate identity from generic prompts and category clichés.
That usually produces:
- safe type combinations;
- predictable gradients or neutral palettes;
- recycled layout patterns;
- copy that sounds politely interchangeable.
The result may be presentable. It is rarely ownable.
The role of constraints
Good brand systems are constraints engines.
They tell the production process what must stay true:
- how the brand speaks;
- what visual tensions it keeps;
- what category codes it uses or rejects;
- what the work must never drift into.
AI becomes much more useful when those constraints exist first. Without them, the model falls back to statistical middle-of-the-road answers.
Analysis: the stronger the constraints, the less likely the brand is to collapse into a category average.
Why this affects discovery
Distinctiveness is not only a design concern. It affects discoverability too.
If a brand looks and sounds like every other AI-assisted site in the category, then screenshots, previews, visual search surfaces, and remembered impressions all get weaker. The page may be readable but not memorable.
CTA: If you use AI in branding or web production, do not ask it to invent your identity. Ask it to work inside a real one.
What a real constraint looks like
"Add stronger constraints" is easy to nod at and hard to act on, so it helps to see the difference between a weak constraint and one that actually bends the output.
A weak constraint is a mood word. "Modern, clean, trustworthy" constrains nothing, because every option in the model's training data already claims to be all three. Feed that to a system and you get the category average back, which is exactly the sameness the backlash is reacting to.
A real constraint closes doors. "No gradients. One accent colour, used only for actions. Headlines set in a serif with visible history, never a geometric sans. Copy that names the client's actual problem in the first line, never a mission statement." Each of those rules removes a large region of the default space, and what is left is narrower, stranger, and far more ownable. The test is simple: if a constraint does not forbid something a competitor would happily do, it is not a constraint, it is a preference.
This is why constraints have to come from strategy, not from the tool. A model cannot tell you which doors to close, because closing them is a business decision about who you are and who you are not for. Give it those decisions and AI becomes a fast way to explore inside a distinctive box. Withhold them and it defaults to the middle, every time, no matter how good the prompt.