Why Does All AI Art Look Like That?

Way back in the 1990s, I went to Art School. It was challenging, fun, and informative and among many of the memories I have from the time, one stands out with regard to the ongoing discussion about AI-generated art.

Life Drawing was one of the first classes all first-semester students had to take. This is the class where you spend 2 hours a day meticulously studying a bowl of fruit or box and pole, or a naked person in a curious pose. These life-drawing models would be set up and lit by harsh lamps to cast dramatic shadows, so we could observe and learn how light and atmosphere effects surface texture and creates the composition. It’s a challenging class because it is A) extremely technical and B) extremely boring.

It didn’t matter what your interest or major was — painting, sculpture, photography, interior design, Fashion (or, like me), sequential art — no one likes drawing sweaty naked dudes sitting on a box for two hours. But it is important because this is how you learn the fundamentals of all art, regardless of medium.

One student in particular, had trouble in life drawing. He was, like me, interested in sequential art. He wanted to draw comics for Marvel and didn’t really get why he needed to spend four hours a week drawing the sweaty naked dude sitting on a box instead.

The first week of class, the instructor had us bring a piece of our art. Something we were proud of, to show the class. It was both an icebreaker and a way to introduce the class to the idea of public critique.

My fellow comic book artist showed us all a page from his comic book portfolio that he was very proud of, and which he had very obviously traced from one or more published comics. The figures were stiff, familiar, but different, as he’d slathered on his own details and changed just enough of the figure to make them “original characters” while still leaving enough detail underneath for those of us familiar with the recent Marvel and DC publication schedule to identify which issue and page he’d cribbed from. The backgrounds were minimal to non-existent, and the perspective was likewise barely there.

I felt embarrassed on his behalf. Not because this childish attempt at art was so over the top and unpolished but because it was so familiar. That’s how many of us who were interested in comics started. We’d trace our favorite panels and draw our favorite characters, doodle battles, try our hand at layouts; fail, and try again.

And this is why we were there at art school, to learn how to do this. Our professor very patiently and kindly ripped this student a new one. He was firm but nice about it, but clearly, it left an impression on my fellow aspiring comic book artist. This was probably the first time anyone had told him his art sucked. And even though they did it with aplomb and in the gentlest of ways, it’s never fun to hear that what you thought was good work, was just a derivative mess.

Thinking back on it now, that fellow classmate then would have loved AI-generated art, because it did exactly what he wanted: it childishly mimicked the real thing in his head. Certainly, a few contemporary AI “artists” have never experienced any level of critique, until they posted their “work” on social media.

By now you’ve seen some choice examples of AI art (or possibly even created some — no judgment, this is a safe space). Also by now, you’ve probably asked yourself, “why does it all look *that* way?”

You may not be able to identify the particulars of what *that* is, but you’ve seen enough of it to know at a glance at some thumbnail on Twitter or Facebook to go, “right, That was made using Midjourney.”

Since this is a topic of discussion, one I actually have some knowledge of, I thought I’d take a moment to outline the 6 ways to identify AI-generated art (and why they look like… that):

1. Uniform Texture. Hair, skin, necklace, clothing — all of it looks like it was made from injection-molded plastic. Human artists spend years learning techniques to render different materials so they have varying textures. The algorithm behind AI-generated art uh, does not do texture very well. At all. It’s always too smooth, or too busy, with no sense or actual thought behind where the details go. There’s no foreshortening or blurring, no focal point. The Algorithm has no idea where the eye of the viewer should be drawn and so just moves it everywhere, all at once.

2. Incompetent Anatomy. AI famously doesn’t know how many fingers a human has, but there are also other cues when you can’t see hands. It’s a little harder to tell with a piece that’s mimicking stylized cartoonish anatomy (anime, for example) but the misproportioned eyes and lopsided body construction are a dead giveaway. Also three rows of teeth (why is it always three rows?) AI art doesn’t build on an armature of sketched-out anatomical building blocks checked for composition and relationship, to ensure the figure is dynamic but based on some anatomical model resembling a human (or anything). There’s no underpainting. No trial and error. That takes time. And speed is the essence of AI art.

3. Lack of Perspective. AI art lacks pers[ective (and I don’t just mean the ability to step back and engage in critical self-analysis, though that’s an issue among the proponents of AI art). What I mean by lack of perspective is, AI art has no concept of fore-middle-background, so everything looks flat. There’s no dimensionality, no atmosphere, no foreshortening. Everything is rendered in soft detail, like you’re not really focusing your eyes, just glancing at shapes in space.

4. Flat Camera Angle. This one is a bit trickier to notice if you’re a novice or someone not versed in art, especially cinematography, but once you do see it, it’s all you can see: AI images are always eye-level head-on, what’s referred to in TV and movie making as American plan. It’s composed as if you the viewer are sitting in the space with the subject, looking at it from arm’s length, with the focus at eye level. AI-generated art can’t do birds-eye view, worms-eye view, 3/4 high/low, Dutch angles — nada.

5. Indistinguishable lighting. Where is the light source in this portrait? Who the hell knows, because there’s not enough contrast in the shading to tell. Even when there is heavy shadow/light, AI images have puddles of light from random sources. The Algorithm is incapable of considering dimensionality, and where a light source might be coming from, especially if that light source isn’t in the frame of the actual piece. Everything looks like it’s shot on an overcast day, or under soft, studio fluorescent light. Or worse, it has a bad case of vaseline lens, and everything looks haloed in diffuse light from nowhere.

6. Muddled details. Hair that blends into a necklace; superfluous limbs; background figures that emerge from nowhere: these are just some of the classic tells of AI art, where the details just kinda get smooshed together. The reason is, AI can’t handle tangencies (when different objects line up visually so they look like they share a surface). So it does what any inexperienced artist does and fudges it. Obscures it in shadow or smudges over it with greeble (that noodly stuff that looks like an object covered in spaghetti or melted wax).

Proponents of AI art have argued that it’s still the early days, and they’ll get these limitations sorted out. Maybe they will, maybe they won’t. But someone should tell the eager beavers who are applying for jobs to replace half a dozen concept artists to slow their roll and maybe put a little more time into learning the fundamentals of why smears of pigment and graphite and chalk on paper or board can convince us it’s a picture of a person or a mountain or a sunset.

Just because Bob Ross has given you permission to be generous with the happy little trees, it doesn’t mean you can make a living as a landscape painter.

After four years in art school, I learned a lot. So did that hapless classmate from Life Drawing back in freshman year. As we were in the same department, which wasn’t that big at the time, I ran into him from time to time. During our senior year projects, I went to a viewing of work by soon-to-graduate students and he had a couple of pages of comics up. The figures were vibrant, and dynamic, crafted with a skill that he had gained from four years of hard work and determination. He took our kind professor’s advice to heart and learned how to draw a figure in space so that it had volume, texture, perspective — life.

Maybe AI Art will get to the point one day where it can make convincing real art. I doubt it, but if it does happen, it will because someone, somewhere took the time to teach it how to make art, not just trace, copy, and manipulate someone else’s.