AI Image Generation Has Gotten "Better," But Its Reputation is Only Getting Worse

The problem was never early-stage sloppiness; it was the method itself.

“AI art may be inferior to traditional art now, but just wait!”

That’s been the refrain of AI art defenders for years. Indeed, thanks to great improvements in the models, the early complaints of weird hands and sloppy text no longer apply. Anyone can type just about anything (appropriate) into a free image generator and expect a nearly photorealistic picture faithful to the prompt with no glaring errors, a fresh custom creation the world had never seen before.

But we don’t hear early critics saying, “You were right - it did get better. This truly is a miracle.” If anything, they’ve only gotten more critical.

To defenders, the skeptics just sound like the Luddites who didn’t want to lose their jobs or the “experts” who thought the Internet would be a passing fad. It’s true that many groundbreaking innovations faced early pushback that looks foolish in hindsight…but it’s not true that early pushback is a sign that an invention will join that list. There are better analogies to be made here.

The microwave oven was hailed as cutting-edge in its early days, praised for its incredible convenience. New cookbooks popped up with microwave recipes for everything from cake to steak. But even as the tech has gotten better and the devices have become household staples, we still uphold good old methods of pan-frying, roasting, and baking as the standard for proper cooking. We accept that microwaving is a shortcut that sacrifices quality. Convenience is its only upside.

So too can we accept that a conveniently generated AI image is like a meal whipped up in the microwave. Just by nature of the method, we expect that the texture won’t be ideal and that it’ll be unevenly cooked. It can be passable, but “passable” is the best we can hope for. It will never be a superior way, and it’s not dogged traditionalism that makes us say this; it’s the results themselves.

There’s another source of inferior substitutes that unfortunately surround us every day: plastic. Early plastic proponents said this would be the do-all material of the future. Compared to its predecessor of ceramic, glass, wood, or metal, a product’s plastic version is less reliable and less tasteful-looking, but it's much cheaper. The more ubiquitous it would become, the more it would set the standard for what the good should look like, how well it should function, and how much it should cost.

Serious criticism wouldn’t be heard until decades later, when plastic was found clogging up water sources. Knowing what we know now, we wouldn’t have wanted all this plastic in the first place, but we can’t easily go back. Given how quickly AI images polluted the Internet, we didn’t have to wait that long.

AI-generated images sure seem to have a microwaved, plasticky quality to them. It's time for AI art fans to accept that the public now views them the same way. It seems like statistical pastiche, even at its best, is just not a good method for creating art.

Why make that blanket judgment when there's so much variability in AI art and it keeps improving? Proponents can cherry-pick examples of “good” AI art alongside examples of “bad” human-created art and insist we can't say AI art is inherently any worse. I'm not going to insist there's some ineffable “soul” detectable in all human-created art and claim all AI art is easily clockable, but to deny the problem is playing dumb. There’s a new art style sweeping Facebook, Google Images, YouTube, Amazon products - and people don’t like it. The term “slop” wasn’t coined for no reason.

Now that the models have gotten “good” enough, it’s clear that what makes their output “slop” isn’t early-stage technical sloppiness; it’s the method itself. One area in which AI systems cannot beat the human mind is few-shot learning. We don’t need thousands upon thousands of examples to understand why not to draw weird limbs, inappropriate textures, nonsensical backgrounds, incoherent details. A person with a real inner world has a unified vision of what they’re trying to achieve and how it’ll likely come across to real human viewers. If they don’t have the artistic chops to pull off realism, they can still create effective illustrations through a coherent cartoon style, “boiling a visual down to its pure essence through process…of elimination.” Only once AI arrived on the scene did we start to see images with both (once-)impressive technical rendering and needless errors revealing that no human vision was involved. The errors are getting smaller and subtler, but no less jarring.

It would be one thing if AI image generators were making the Internet a more beautiful place and illustrators were still complaining that it would take their jobs. But when the demand for illustrators is drying up thanks to a process that visibly isn’t any better, they have every right to complain - and so do we consumers.

AI art pushers accuse critics of wanting to “gatekeep” art. Turns out, putting imagery on the Web is a process that deserves to be gatekept. If a picture isn’t a slice of reality or a product of human ability, there really is no need for it. No matter how much the models improve, the public now understands that at worst, they empower spammers and tricksters, and at best, they’re simply irrelevant.