So Substack offers an AI image generator for our posts. Cool! I have actually used it for a quick, weird, copyright-free image for a post. I have always been fascinated by the aesthetics of AI images and what the AI chooses to add that was not requested, which we’ll explore for a bit below.
This all started because I entered the prompt “man under a lake” and I could NOT, no matter what I entered, get it to show a man under water. Let’s try to get it to create an image of a man under water and see what happens. PS, I am using the “Neon” setting (as opposed to photo, etching, cartoon, etc.) The Photo setting DID generate a picture of a man underwater, for instance, although it was still plenty wierd!
The above image was generated from the prompt “Man under a lake.” Each prompt generates four images, none of which featured a man under water. One of them had a man standing next to a car. In the image above, I did not ask for the moon (I don’t ask for the moon!) or the island. The aesthetic is kind of ‘80s retrowave in color choice and emphasis on lines.
This one was generated by the prompt “Man under water.” You will notice that he is not UNDER water. I’m curious that in all four images there was a silhouette of a man facing away, which is kind of a movie poster image, speaking to how AI is trained on existing media. I did not ask for the city, nor did I ask for it to be so explicitly based after Blade Runner. This illuminates a very important point: It’s an illusion to believe that AI is a reflection of our world, when it is actually a reflection of our media.
AI models are trained by sucking up media, so they reflect a world that has already been parsed and processed, not the actual world in all its messy, non-data form.
In continuing to try to get an image of a man under water, the prompt for the image above was: “man under under under under under under under water.” Apparently repetition does not work for emphasis. You’ll notice that I mentioned nothing about cars, or cars in city apartments, or really anything related to the image above.
I thought maybe the issue might be that I’m asking for an image of a man? So I used the prompt “Woman underwater.” I’m curious that the woman generated is white, thin, attractive, in revealing clothing and apparently in full makeup. Surely you recall the dustup over Google’s AI engine and how users could NOT get it to display a white person, even when they requested German soldiers or whatnot. There may be some work to be done to reflect diversity in AI images, but it seems that the AI systems themselves are not going to do it! BTW, ALL the women generated looked a bit like Scarlett Johansson (and all were thin and magazine-style attractive), so we have to wonder about that as well.
I thought perhaps I could get an image of a city underwater? But no dice. The image is pleasant to look at, however. Notice how all the buildings are basically the same; no variations in architecture here! I just threw it in because golly, it’s pretty! I think I once saw a guy in Times Square making images like this with spray paint.
Okay, I’m starting to get bored, so I thought if I just ask for “sea creatures in the ocean” I might get some underwater images. Two of the four returned were under the surface of water. All of them returned nonsensical, deformed sea creatures. This one was special by having the surface of the ocean UNDER the ocean, which you might admit is a little bit extra on the part of the AI.
I’ll leave it up to you to determine what it all means, but I hope you find it as fascinating as I did. Needless to say, I don’t think I would trust our current AI to perform brain surgery… or order a chicken bacon wrap. But it does create images that look like something, so yeah. It does that.