Maybe AI Sucks, Or Maybe People Just Suck At AI

It takes practice, just like anything else. And understanding limits.

One of my favorite writers on ye olde interwebs is Freddie deBoer. I find myself agreeing with most of his takes and he’s always made me want to be a better writer.

But I do find myself disagreeing with Freddie on his latest take, which feels a bit like a strawman critique of AI. Don’t get me wrong, I am not a futurist. I don’t believe all the hype (or doom and gloom) surrounding AI artwork. I agree with virtually every critique leveled at AI writing: that it is soulless and hollow and, worst of all, boring.

I’m more interested in AI artwork, however. I’m not an artist (or at least not a very good one) so having AI tools has been a fun diversion and hobby of mine over the past few months, and I think I’ve gotten pretty good at it, though others are far better. Like anything else, AI is a tool. You can use it by itself, or combine it with other programs like Photoshop, or integrate it into physical media.

Freddie uses two examples of the shortcomings of AI art which I think misrepresent, to some degree, the actual limitations of a program like Midjourney. The first is this side-by-side image of an AI rendition of actor John Candy compared to a traditional cartoon version of the late actor:

I agree that the AI image of John Candy is terrible, but that’s because whoever was making the image was bad at it (or unlucky in their choice of subject; some actors are impossible—try Brian Cox sometime). But I can make an image of John Candy that I think is as good as the cartoon image above:

Or another:

Freddie’s second example is this terrible approximation of actress Goldie Hawn:

Midjourney is pretty bad at making some actor likenesses for whatever reason. I’ve found that women, oddly enough, are harder to get right. Realism, however, is also part of the problem. We run headlong into the Uncanny Valley when we start to make “realistic” versions of these famous people, just like we do when we make CGI versions. For instance, the ghoulish rendition of Grand Moff Tarkin in Rogue One:

But AI is still pretty good at making stylistic versions of famous people. Here are a couple Goldie Hawn portraits I made for this post:

And in living color:

These look like Goldie Hawn to me. They’re also kind of cool images! I’m not sure you could easily identify them as AI if you weren’t told ahead of time, because I’m not trying to make super realistic stuff here.

You can, however, do realistic if you know the tricks. I’ve shared my Seven Demons idea in the past—a fictional film or TV series about time-traveling heroes who go to various times and places, such as feudal Japan and the Wild West, and fight against time-traveling villains that’s based on Seven Samurai and The Magnificent Seven—which I made a bunch of (I think) pretty cool AI images for. Here are some:

Steven Yeun

Mads Mikkelsen

Bryan Cranston

J.K. Simmons

In any case, I think it’s obvious that a great range of quality and skill exists in the world of AI artwork / art generation, whatever you want to call it. I find that making these images black and white helped them appear more real—and indeed, many people Googled ‘Seven Demons’ when I shared these, because many people thought it was a real film. Another trick I used was using branding on many of the images I shared, like so:

I did something similar when I came up with a bunch of fake Grand Theft Auto VI images:

I don’t think these suck. What was difficult, however, was creating images of women that worked in my Seven Demons motif. I really wanted to include Michelle Yeoh, for instance, but I couldn’t get a single image to look like her—at least not her at her current age. I’m not sure why this is, but I have a hunch AI is just sexist. Which, yeah, that sucks. Maybe AI does suck.