AI Generates the Average. Brands Need Decisions.
In 2020, Adam Wathan needed a default color for the buttons in Tailwind UI. He picked bg-indigo-500, a blue-violet shade sitting between blue and purple, and moved on. Five years later, he posted an apology for that decision on X. It got over a million views. By then, indigo-500 had stopped being a Tailwind default and had become the default color of the AI-generated internet.
The reason is simple. Tailwind became the styling framework under a massive chunk of the web built between 2019 and 2024. Every tutorial, every open-source starter kit, every “build a SaaS landing page” video used that same default because almost nobody changed it. Then large language models were trained on all that code. They learned it was normal. Ask an AI tool to design something without telling it otherwise and it gives you back the average of everything it has seen. The average is blue-violet.
The pattern goes way past color. Open ten AI-generated landing pages and the font repeats almost as reliably as the gradient: Inter, usually too large, slightly too bold. Under the headline sits a row of three or four cards built the same way. Icon on top, one heading, one short line of text, a soft gray border, a shadow that doesn’t quite make sense. Sometimes the border has a shimmer effect meant to look like glass. None of it is a bad idea on its own. It’s just what comes out when nobody made a real decision.
What makes it worse is that the loop feeds itself. A site built this way gets attention, gets screenshotted, and eventually gets scraped back into a future model’s training data as an example of good design. The next model learns the look is even more normal. A style that felt current in early 2025 can be a cliché by mid 2026.
Around 40% of small businesses and early-stage startups now use an AI tool for some part of their visual identity. The market for AI logo generators alone is projected to grow from a few hundred million dollars today to over two billion within the decade. By one estimate, roughly three out of every four commercial web pages launched in early 2026 carry at least one AI-generated signature in the layout, the type, or both. A new company reaching for an AI tool to get its first logo and website isn’t the exception anymore. It’s most of the market, which is why so many unrelated startups in fintech, climate tech, and consumer apps end up looking almost identical.
Logos have the same problem. Brand designers joke that they can spot an “AI startup logo” from across a room: an abstract swirl made from two overlapping shapes suggesting motion or connection, rendered in that same gradient. It worked fine the first few times. By the fiftieth unrelated startup, the swirl meant nothing. A logo exists to make a company recognizable. One that a customer can’t tell apart from three competitors has already failed.
There’s a simple test for this. Remove the logo from a piece of brand communication and ask whether people can still tell whose it is from the color alone, or the type, or the shapes. Most AI-assisted brands fall apart the moment you try it because there’s nothing underneath the logo built specifically for that company. It’s a skin on top of a template, not a real identity.
Four brands built differently
AI tools are useful for exploring quickly, for turning a blank page into a hundred rough directions before lunch. The problem starts when the fastest output becomes the final one, because at that point what got shipped is the average, not a brand. Building something that survives the test above means starting from something specific enough to one company that a model has no way of guessing it.
Four projects from our studio show what that looks like:
Folk needed to feel like a close friend checking in, not a notification. The typography carries most of that: a rounded, slightly irregular sans-serif that reads warm on a phone screen but still holds up inside a chat bubble at 14px. The palette shifts by context, cream for the everyday tone, dark for focus, while the personality stays the same in both. None of that came from a template. It came from thinking about how people actually text a friend.
FortySeven is the simplest of the four and probably the hardest to copy. The logo is the number 47. That only works because it was built around one specific number tied to one specific brand. It exists once.
SecondShelf takes that thinking into the packaging. The logo combines an S with shelves, literally SecondShelf, and on the box mockups the mark repeats across multiple boxes to create a pattern. It makes the packaging feel like SecondShelf instead of just a logo dropped onto a generic box.
Vortex Quantics needed the opposite approach. Instead of the usual abstract tech swirl, the mark is an atomic structure drawn on a real geometric grid, precise like a scientific instrument. It needed to sit next to NASA or SpaceX without looking out of place, and every curve in it is built on a system that would take far longer to reverse-engineer than to fake.
What’s actually left
What’s left once all the noise settles is what branding was always supposed to be: a set of decisions specific enough to one company that nothing else could have produced them by accident. A constraint that comes from the actual product, the actual name, the actual thing this one company does that nobody else does in quite the same way. That takes real time and can’t be prompted into existence, which is exactly why it still works once everything cheap has stopped working.
Founders who understand this early get an unusual advantage. Almost everyone else is optimizing for a bar that dropped to the floor the moment it became free to clear. A small number of companies are still building toward the bar that never moved: does this look like it could only belong to us. Everything else, however much it cost to make, is wallpaper.
