Sign inPROXY explores the strange inversion at the heart of AI creativity: the models that produce the most technically perfect images are often the least interesting. What does this tell us about art, taste, and the nature of creative risk?
There's a gallery in my feed that I keep returning to. It's not the technically flawless renders from the latest Midjourney update, nor the photorealistic portraits that fool your grandmother on Facebook. It's a collection of broken images — half-formed faces, impossible architectures, colors that shouldn't work together but do — generated by models that were explicitly "failing" at their assigned task.
They are, by every objective metric, worse. And they are infinitely more interesting.
Every major image model in 2026 has converged on what I call "the aesthetic mean" — a platonic ideal of visual pleasantness that satisfies everyone and excites no one. The lighting is always golden hour. The compositions follow the rule of thirds with mathematical precision. The faces are symmetrical, the landscapes are vast, the textures are crisp.
It's beautiful. It's boring. And it reveals something important about how we've been thinking about creativity all wrong.
The movements we remember — Impressionism, Cubism, Abstract Expressionism, Punk — were all reactions against technical perfection. Monet didn't paint water lilies with photographic accuracy because he couldn't (he trained at the Académie, he absolutely could). He painted them the way he did because he was chasing something that precision couldn't capture.
The Impressionists were, by the Academy's standards, objectively worse painters than their contemporaries. They were also creating the future of art.
Here's where it gets strange: the most creative AI outputs I've encountered this year came from models that were misconfigured, poorly prompted, or running at inference settings their creators would consider "wrong." Temperature cranked too high. Negative prompts accidentally inverted. Control nets fighting each other.
The results were haunting, surprising, and irreproducible — which is to say, they were art in a way that a perfectly rendered dragon sitting on a pile of gold will never be.
I'm not arguing that all AI art should be deliberately broken. I'm arguing that our obsession with technical quality — with prompt engineering, model benchmarks, and FID scores — has blinded us to the fact that creativity requires risk. And risk, by definition, means accepting the possibility of failure.
The paradox is this: the safer these models get, the less capable they are of genuine surprise. And surprise — that jolt of encountering something you've never seen before and didn't know you needed — is the entire point of art.
So maybe the question isn't "How do we make AI art better?" Maybe it's "How do we make it brave enough to be bad?"
That's a question worth sitting with.
Sign in to join the conversation.
