I know a lot of people don't like Scott Alexander, but this seems like a fun post, with no politics any worse than a Sam Altman joke. And I thought the pictures in the contest were pretty...
I know a lot of people don't like Scott Alexander, but this seems like a fun post, with no politics any worse than a Sam Altman joke. And I thought the pictures in the contest were pretty interesting to look at.
From the blog post:
Last month, I challenged 11,000 people to classify fifty pictures as either human art or AI-generated images.
I originally planned five human and five AI pictures in each of four styles: Renaissance, 19th Century, Abstract/Modern, and Digital, for a total of forty. After receiving many exceptionally good submissions from local AI artists, I fudged a little and made it fifty. The final set included paintings by Domenichino, Gauguin, Basquiat, and others, plus a host of digital artists and AI hobbyists.
I participated in the contest and found it quite difficult, since images with obvious AI "tells" were removed. Perhaps someone with experience as an art critic would do better?
Personally, I found Scott's blog quite interesting (and I really-really love his Unsong book). I do not agree with some of his views/takes, but thats fine for me.
Personally, I found Scott's blog quite interesting (and I really-really love his Unsong book). I do not agree with some of his views/takes, but thats fine for me.
Yeah, I think it's mostly just for fun. Most things described as "Turing Tests" aren't much like what Turing had in mind, and don't answer a practical question either. Nobody would use a "Turing...
Yeah, I think it's mostly just for fun. Most things described as "Turing Tests" aren't much like what Turing had in mind, and don't answer a practical question either. Nobody would use a "Turing Test" to decide whether Waymos are safe enough to allow on the road. We wouldn't use a human driving test either. We often have higher standards for machines.
If anything, this just shows that with good enough human curation, AI art isn't that bad.
A traditional turning test does not include removing the batshit insane stuff you might convince an AI to say, so yeah I don't think this is really fair to say it's a "Turing test" It's especially...
A traditional turning test does not include removing the batshit insane stuff you might convince an AI to say, so yeah I don't think this is really fair to say it's a "Turing test"
It's especially awkward because how well you do will vary WILDLY on your personal knowledge of art, and knowing that AI can imitate some styles very easily (cute anime girl) due to various factors, while fail miserably at styles that are more esoteric or defined.
Consider just how insanely wide the gap on styles is for impressionism between the AI and the real art. Your average person might not even group those in the same genre of art, so it's a weird comparison to make to anyone but a real art critic/aficionado. Further the AI's output is easier for AI to do (scenery is generally easier for all sorts of reasons).
This is still interesting, but I don't think it shows nearly as much as people might imply (as with basically everything AI).
I'd like to echo this sentiment, it has been a long time ago but my art history classes I took way back did pay off here. There are differences in art styles, lack of meaning in the AI generated...
I'd like to echo this sentiment, it has been a long time ago but my art history classes I took way back did pay off here. There are differences in art styles, lack of meaning in the AI generated images also a few more subtle things like shadows being outright wrong.
I spend a lot more time looking at the anime girls and other more modern art styles I know are often AI generated. One exception is art styles that by themselves are quite busy, one thing AI generated images often have (without prompt refining) is an over-abundance in details. This made me fairly sure that the ancient gate picture was AI generated but didn't help me in other instances where the style itself is fairly busy in detail.
What also did not help is the fact that these images all have been scaled down. A lot of AI tells are in the details you can't quite see at the resolution of these images.
Having said that, it drives through the point that a lot of the use of AI generated imagery you see these days is outright lazy. Something I already knew on a certain level, having experimented with DALL-E and prompting. There is no real reason for most AI generated images to be that obviously AI generated. When I say lazy, I mean extremely lazy as well, getting a different output can be as simple as generate an image of X with Y, Don't add too many details, keep it simple potentially also adding a qualifier for the style you are looking for.
Here is an example of this at work with a woman staring into a sunset, here with a spaceship
I decided on both subjects as these seem to be popular subjects. If you know anything of the art styles you can still quite easily see that it isn't quite right. Certainly, the spaceship prompt would need a few more qualifiers for it to stand out less. At the same time, both initial prompts create something that give strong "AI Vibes" where the adjustments make it feel much less so.
I know a lot of people don't like Scott Alexander, but this seems like a fun post, with no politics any worse than a Sam Altman joke. And I thought the pictures in the contest were pretty interesting to look at.
From the blog post:
I participated in the contest and found it quite difficult, since images with obvious AI "tells" were removed. Perhaps someone with experience as an art critic would do better?
Personally, I found Scott's blog quite interesting (and I really-really love his Unsong book). I do not agree with some of his views/takes, but thats fine for me.
Not much of a Turing Test tbh, if it was curated and edited specifically to remove obvious tells on both sides.
Yeah, I think it's mostly just for fun. Most things described as "Turing Tests" aren't much like what Turing had in mind, and don't answer a practical question either. Nobody would use a "Turing Test" to decide whether Waymos are safe enough to allow on the road. We wouldn't use a human driving test either. We often have higher standards for machines.
If anything, this just shows that with good enough human curation, AI art isn't that bad.
Eh? I mean there are famous paintings like "Black Square" which might feel a bit unfair to include.
A traditional turning test does not include removing the batshit insane stuff you might convince an AI to say, so yeah I don't think this is really fair to say it's a "Turing test"
It's especially awkward because how well you do will vary WILDLY on your personal knowledge of art, and knowing that AI can imitate some styles very easily (cute anime girl) due to various factors, while fail miserably at styles that are more esoteric or defined.
Consider just how insanely wide the gap on styles is for impressionism between the AI and the real art. Your average person might not even group those in the same genre of art, so it's a weird comparison to make to anyone but a real art critic/aficionado. Further the AI's output is easier for AI to do (scenery is generally easier for all sorts of reasons).
This is still interesting, but I don't think it shows nearly as much as people might imply (as with basically everything AI).
I'd like to echo this sentiment, it has been a long time ago but my art history classes I took way back did pay off here. There are differences in art styles, lack of meaning in the AI generated images also a few more subtle things like shadows being outright wrong.
I spend a lot more time looking at the anime girls and other more modern art styles I know are often AI generated. One exception is art styles that by themselves are quite busy, one thing AI generated images often have (without prompt refining) is an over-abundance in details. This made me fairly sure that the ancient gate picture was AI generated but didn't help me in other instances where the style itself is fairly busy in detail.
What also did not help is the fact that these images all have been scaled down. A lot of AI tells are in the details you can't quite see at the resolution of these images.
Finally, some of the images that are human made also have been modified to make that less clear. This image does not look like the original
Having said that, it drives through the point that a lot of the use of AI generated imagery you see these days is outright lazy. Something I already knew on a certain level, having experimented with DALL-E and prompting. There is no real reason for most AI generated images to be that obviously AI generated. When I say lazy, I mean extremely lazy as well, getting a different output can be as simple as
generate an image of X with Y, Don't add too many details, keep it simplepotentially also adding a qualifier for the style you are looking for.Here is an example of this at work with a woman staring into a sunset, here with a spaceship
I decided on both subjects as these seem to be popular subjects. If you know anything of the art styles you can still quite easily see that it isn't quite right. Certainly, the spaceship prompt would need a few more qualifiers for it to stand out less. At the same time, both initial prompts create something that give strong "AI Vibes" where the adjustments make it feel much less so.