I know some of you clicked through to the art experiment over at Astral Codex Ten, where you were asked to see if you could identify human vs AI art. I did the test, in a casual “clicking rapidly through it, not trying especially hard to get it right.” I found it interesting and also was genuinely surprised by some of the AI pieces. I should say that I fall solidly into the cliche “I don’t know anything about art, but I know what I like.”
Here are the results of the test, if you’d like to see how well or badly respondents did overall. Apparently there were 11,000 respondents, give or take, which is a pretty amazing sample size, honestly.
There’s a general discussion, including an interesting and detailed discussion about this image:
Which I liked a lot, but now like less after reading a long description of the many things that are wrong with it. Those things don’t bother me that much, but fine, I admit my opinion has been influenced nevertheless and as I say, now I like it less. I’m glad the giant ship was human and that nobody spent any time explaining what was wrong with it, because I really liked that too, and for a lot of the same reasons: I like monumental architecture (and, apparently, monumental ships) with lots of detail. You can click over to Astral Codex Ten and look at the ship; I’m not posting that image here because it is, after all, by a real person and therefore copyrighted.
A lot of people apparently shared my view that anything impressionistic was probably human, or ought to be human, and therefore had trouble with that category. One image that I liked a lot turned out to be AI generated and also was the single most-liked piece in the whole test. It was this one:
I still like it and I still find it hard to believe it’s AI generated. Some of you (Hi, Elaine’s Teen!) were much better at this than I was. If you see this post, then did you recognize this one as AI generated? And if so, how? There was another one, “Leafy Lane,” that to me seems similar to the one above and which I also liked and also thought was human-generated. Here it is:
Another favorite of mine was this one:
Which I felt was bright and cheerful and I just liked it. I still do.
Here’s one I thought was by a real person:
I don’t like it; to me, the figure of the woman looks like she’s been plonked down on top of a landscape painting without any attempt to make her look like she’s really there. BUT, I could swear I’ve seen plenty of paintings where the human figures looked unreal or misplaced to me in exactly the same way, so I guessed human. (Remember, I took about five seconds to guess one way or the other. But I really did have this exact reaction: she looks wrong, but in a human way. Nope, turns out she’s wrong in an AI way.)
Scott says he tried not to include any with hands that were obviously wrong. I think he missed one. The left hand of “Muscular Man” looks very strange to me.
How can Scott have thought the left hand passed muster? Everything starting partway down the forearm looks just strange.
Everyone, including me, guessed correctly that this next one was AI.
I like this one a lot — something I apparently like is intricate detail — so I’ll end with it.
To see the human artworks and the discussion, click through.
Please Feel Free to Share:
That most-liked picture is very reminiscent of Van Gogh’s Café terrace by night, within a more rural rather than old city-street setting.
The row of trees along the path and garden wall also reminds me of one he painted at Arles IIRC, in a different color-tone but with a similar composition.
Not that that would have been enough to guess it as AI, since another artist might have composed and painted a similar scene.
And looking at the whole pictures instead of the details, I completely missed some things that were very obviously wrong once I saw them mentioned in the comments, like the Esher column in the ruin and the baby’s thumb held by the mother’s hand on the back of his thigh.
River cafe was actually pretty simple:
details weren’t where they ought to be. Details faded too uniformly.
The street didn’t end, or trail off into the distance, it just stopped abruptly. It hadn’t continued then been painted over either. It doesn’t go of as a shadowy lane, or a dark tunnel.
Assorted objects don’t have shadows where logic would put them- take the lamp post in the foreground. If it’s lit, the stone wall next to it should have a bit of a highlight on the side with the lamp. The wall doesn’t, it has a shadow instead.
The lamp farther down the road is apparently lit, as it seems to have a line of yellow light coming down the road from it, but the tree it’s next to doesn’t have any indication of that light.
The other stone wall is lit from the river side, maintaining continuity with the nearer one, but while the walls have shadows on the sides facing us, they aren’t casting shadows onto the street. The tables are uniformly lit as if from above, not by the warm light from the cafe windows or from the lamp and sunrise. The condiment pots on them have highlights but cast no shadows. The shadows of the tables are vast by an overhead source of some kind. Chairs don’t all have shadows. The foreground lamp post doesn’t have any shadow at all, not from any of the assorted potential sources- it ought to be falling either foreground left or foreground right.
The distant sunrise, oddly enough, isn’t a problem. I’ve seen sunrises that look like that and leave the landscape that dark before them- it happens when the sun isn’t even on the horizon yet, but the first fingers of dawn have reached the cloud layers. (Get up at 4:15 every day, you see a lot of sunrises.) But for that sunrise to be reflecting in the water (whether or not a real one would), painters logic would lighten the distant landscape more than it has.
Compare to this:
https://www.ebay.com/itm/356296885775
or this
https://www.ebay.com/itm/356308424079
or this
https://www.ebay.com/itm/356308411404?itmmeta=01JDMP3NP0JSRKMNWB38Y982DV&hash=item52f5a2e40c:g:mqwAAOSwip5nRfs7
In the last one, look at the sky. Follow it as it recedes. Then look at the land, and follow it as it recedes. Can you tell where the characters are looking, by the angle of their heads, the shapes of their bodies?
I’m still very fond of that one. It’s one of my best. Not the greatest photo though. It wasn’t meant to be seen closeup, but it was the only way to keep a reflection of the camera off the picture.
That’s really interesting, Teen. I can go back and forth from your comments to the image and see that you’re right, but absolutely none of that seems obvious to me.
I really like the Phantom Grasses one, by the way.
Hanneke, I think my vague memory of the Van Gogh painting gets in the way. I feel this ought to be real because I have this vague memory of a real painting.
More people than I’d expect like the grasses. I’m still quite fond of it too, but I guess its’ not the one I expect to get attention? Oh well. It’s full title is Phantom Grasses at Sunset, and I am selling copies. That ebay link is active.
Thanks. Prints are for sale if you’re interested.
Regarding the AI generated image, it takes me some time studying it to mark each of those details and be able to list them. But even without studying them, I see them, and they add to teh level of unreality that says ‘this was AI generated.’