Adapted from something I posted to my Instagram story in February 2025.

Until a few years ago, I assumed a human-level AI designer would show up in the near future. Lately, though, I’ve been drifting toward a more skeptical view: that a genuinely satisfying design AI is not going to appear anytime soon.

There are three main reasons, and most of them have less to do with the limits of the technology itself than with factors outside it.

1. A Small Market

The design market is small. Smaller than you’d ever imagine. The global design market is estimated at around $60 billion. The AI market, which has only just begun to take off, is estimated at around $170 billion. In other words, this newborn AI market is already more than twice the size of the design market. If you were an investor, where would you put your hard-earned money?

DeepSeek made headlines for training a decent-performing LLM on a relatively modest budget, but training AI still demands an astronomical amount of capital. With that kind of money, what AI would you build? What market would you develop your technology for? The LLM market, with an estimated annual growth rate of 33.2%? Or the design market, growing at a mere 5% a year? The answer is obvious.

2. The Gap Between Data and the Real World

GPT, one of the hottest AI models right now, is trained multimodally on an enormous amount of text data and a comparatively smaller amount of image data. My own theory is that LLMs work so well precisely because text data is both abundant and easy to obtain. Image data, by contrast, is enormous in sheer file size compared to text—and video is in a league of its own. That’s why GPT can solve text-based problems remarkably well while still being clumsy when it comes to handling images. (This is purely my own guess.)

But design relies on vision, and vision is intimately tied to the physical world. No matter how well GPT learns from all the data that currently exists, its understanding of the physical world remains superficial. To put it figuratively, it’s like a person who has studied the world entirely from behind a desk. That’s exactly why the big tech companies are now scrambling to build robots.

Humans, they say, can perceive up to about 60 frames per second. People who use gaming monitors might object, but let’s go with it for the sake of argument. If you sleep eight hours a day, you’re awake for sixteen—and during those sixteen hours, we are constantly “seeing.” By that math, we take in 3,456,000 images in a single day.

Add up all the images you’ve seen over a lifetime and you get a staggering number. And the data we see is even larger in terms of sheer volume. We see the world across an enormously wide dynamic range that goes well beyond 0 to 255. Not all of it makes for high-quality training data, of course, but measured in count or in file size it amounts to an astronomical figure. It’s simply no contest against GPT, which can only see photographs taken by a camera, confined to the RGB range of 0 to 255.

3. The Points Where Interpolation Falls Short

Modern AI models are excellent at finding patterns within given data and interpolating between them. Image-generation AI and design AI alike produce new outputs based on previously learned data, but at heart this is closer to a variation or recombination of existing data. Design work, however, is full of creative challenges that can’t be solved by interpolation alone.

That said, perhaps creativity, in a broad sense, is also a form of interpolation. Every piece of research has its related work. Truly novel things are exceedingly rare. There’s a reason for the saying about standing on the shoulders of giants and seeing further. Even so, at least with the AI we’ve had until now, it’s been hard to see interpolation as good as what a human designer can do.