experimenting with the AI, trying to break it as a technique for meta commentary

[tremault]

I was thinking, maybe I try to break it or use it in a subversive way. I think this is using it to make a statement that works as meta commentary and as a means of expression. Anyway, I came upon a prompt that resulted in some funky images. This one best fit my purpose.

The prompt : a lovely dog. his face is rippling and the algorithm is hallucinating. within the algorithm is hidden pictures of dogs and parrots.
My purpose is to subvert it. Or make commentary on it, based on influences listed below. If you want to join in, please try to keep it within the same arena, to break it or make commentary or twist it around somehow.

Here's another one of a man. same deal as the one above.

Edit : influences .
1. https://www.sciencealert.com/these-trippy-images-show-how-google-s-ai-sees-the-world-read-more
2. https://www.ibm.com/topics/ai-hallucinations

 

[unfair]

I consider a lot of AI output kind of broken, that's part of the skill - prompting it to generate what you're actually looking for and sifting through the detritus for the gems that are coherent.

Here are a few of the interesting ones from my prompts in the other thread:

AI turned the bat signal into some nightmare creature:

https://imgur.com/I0MOT3E

How do capes work?

https://imgur.com/e5AkEbC

Early attempts at negating part of the image:

https://imgur.com/Rpyy4fc

Notoriously bad with details like fingers, arrow continuity and direction, bow strings:

https://imgur.com/AimxIJv

Well... that is not the Red Queen I was looking for...

https://imgur.com/dWsqM5T

People playing board games is just generally nightmare fuel:

https://imgur.com/QRXl8fO

https://imgur.com/l0aCUlz

Ah yes... the Wind River portal from hell, with nondescript creature like blobs and an... interesting... idea of how items should be scaled:

https://imgur.com/9d8D3tj

Westworld:

https://imgur.com/tOohF1r

https://imgur.com/ZiTINrv

 

[tremault]

Wow yeah those are broken, but please do try to stick with the topic, it's not about just posting some funny results, it's about purposefully creating subvertive images, or making commentary on the nature of the beast. I would like a pure focus here please.

 

[WilliamRedRobin]

AI's biggest weakness is that it can take in a bunch of images and look at what they have in common, but it can't make a judgement of whether something looks "good" or "bad". I think this is why it struggles so much with things like hands which have lots of natural resting positions. The AI doesn't know what shape a hand is three-dimensionally, so it ends up with odd errors like six fingers when holding a gun because some images have the index finger on the trigger and some have it on the handle.
Ambiguous wording might trick it into combining two unrelated objects, but with people you usually end up with obviously male or female figures because I guess it sees that when people have this feature they're also likely to have that feature, and most of the people it's seen will be models and actors which is why they always seem to have strong jaws and bright eyes.
Combining two concepts that don't usually go together will also cause some confusion. Since there aren't many images of those two things together, the AI is forced to draw from images of the two things individually, and then when it combines them it gets confused about what characteristics belong to which object.

So, one thing an AI might struggle to generate is:

"An ugly, androgenous person with big hands has a bow and cards. There is text."

"A plain-looking, androgenous person with big hands has a bow and cards and is writing. symmetrical."

"A plain-looking, androgenous person with big hands has a bow and cards and is writing."

"A plain-looking, androgenous person with big hands has a bow and cards. There is writing on the wall. Realistic."

​

Something interesting I learned from generating these is that since the algorithm has no way of judging attractiveness, and general politeness prevents us from labelling pictures of people as "ugly" or "plain-looking" it doesn't really know what these words mean. It interprets "ugly" as a monster and "plain-looking" as flat colours, I guess? Maybe there's some societal commentary in there.
It seemed to be ok with the concept of androgyny most of the time. I'd say 6/13 of these figures are truly androgenous and another three have androgynous features, but most of them still have the supermodel facial structure.
It invariably interpreted "bow" as the weapon and "cards" as playing cards, which make sense, since that's what shows up when you google those words, but I was expecting at least some hair bows to show up. "Little girl with a bow" definitely prefers that:

The hands came off surprisingly well, that's probably something the designers of these algorithms are focused on improving since it's such a famous issue, but the bows and cards still have a lot of jank if you look closely. In one image, it's combined a quiver with a pot of pencils which is pretty cute.

That's about all I could glean from this. It's a pretty fun exercise.

 

[Duragizer]

I stopped making pure AI "art" some while ago. I pretty much now only use AI-generated imagery as components in my neo-Dadaist collages. Most of the time, I'm not directing the generator to make janky imagery; I just make use of the happy little accidents I'm provided. But there've been a few times when I've invested meaning in the madness to make a point or elicit a response.

The Lamb Sauce's Gone Bad

Make Anti-Art, Not War

It's Pretty, but Is It Art?

 

[DigModiFicaTion]

I'm not understanding the OP. Are your intended prompts meant as a protest with the hope of breaking down the Ai algorithms? I'm trying to understand the intense adverse feelings toward this tool.

 

[tremault]

I'm not understanding the OP. Are your intended prompts meant as a protest with the hope of breaking down the Ai algorithms? I'm trying to understand the intense adverse feelings toward this tool.

We can't break the algorithm for anyone else, but to know a beast, we must know it's limitations. I prefer to push these tools to see where their points of failure lie. I can easily find the limitations of large language models when it comes to the chat bots, as I have instructed one to give me a certain piece of information without including unwanted information (such as a spoiler) and it is incapable of performing that level of reasoning. That makes it useless for the purposes of safe information retrieval. There is probably similar issues with these image generators, unforeseen issues that could arise further down the line. It's still a major conversation and lots of issues and avenues to explore, I feel that' it's potentially risky for people to give in and trust these tools so readily, as a sort of short cut. It should be approached with caution, rather than like a toy. To me, this feels like a sort of necromancy, and it needs to be challenged and my approach is that of an artist. The whole intent of this topic is as an art experiment. a continuation of the conversation, as art is meant to be. In this way, the soulless chimera or mimic, can be used as a tool to create something that could actually be considered art, as it is being used purposefully with a direct focus on the nature of the thing and as a statement upon that tool itself.
It's hard to communicate though,

 

[DigModiFicaTion]

We can't break the algorithm for anyone else, but to know a beast, we must know it's limitations. I prefer to push these tools to see where their points of failure lie. I can easily find the limitations of large language models when it comes to the chat bots, as I have instructed one to give me a certain piece of information without including unwanted information (such as a spoiler) and it is incapable of performing that level of reasoning. That makes it useless for the purposes of safe information retrieval. There is probably similar issues with these image generators, unforeseen issues that could arise further down the line. It's still a major conversation and lots of issues and avenues to explore, I feel that' it's potentially risky for people to give in and trust these tools so readily, as a sort of short cut. It should be approached with caution, rather than like a toy. To me, this feels like a sort of necromancy, and it needs to be challenged and my approach is that of an artist. The whole intent of this topic is as an art experiment. a continuation of the conversation, as art is meant to be. In this way, the soulless chimera or mimic, can be used as a tool to create something that could actually be considered art, as it is being used purposefully with a direct focus on the nature of the thing and as a statement upon that tool itself.
It's hard to communicate though,

So you are simply trying to disprove it's utility by finding its limitations? Giving a prompt to an image generator isn't trusting it to think for one's self. Everyone here is an artist. Some are perfectly ok with Ai, so I wouldn't use the stance of being an artist as the frame of reference for this nebulous venture. Aren't we all just mimics of those we admire? Isn't fanediting in a way a mimic or mock of those that have and can by those who don't have and can't? I agree that what you are trying to express here is hard to understand. This doesn't appear to be the stepping stone towards the skynet scenario or similar outcomes.

 

[tremault]

So you are simply trying to disprove it's utility by finding its limitations? Giving a prompt to an image generator isn't trusting it to think for one's self. Everyone here is an artist. Some are perfectly ok with Ai, so I wouldn't use the stance of being an artist as the frame of reference for this nebulous venture. Aren't we all just mimics of those we admire? Isn't fanediting in a way a mimic or mock of those that have and can by those who don't have and can't? I agree that what you are trying to express here is hard to understand. This doesn't appear to be the stepping stone towards the skynet scenario or similar outcomes.

I can't always explain exactly how I feel. We often explore how we feel through art. I'm not disproving anything. This isn't a science paper, It's communication of things that I can't readily put into words. Art is about exploring concepts and pushing against boundaries etc.
I'd argue btw, that not everyone here is necessarily an artist, there is a difference between art and design. There are surely lots of artists here and lots of designers but not everyone who is one is also the other. it's nebulous.

 

[tremault]

Prompt for this one :
the artist pushes against the boundaries and tries to understand the true nature of the mimic that is AI

 

[unfair]

To me, this feels like a sort of necromancy, and it needs to be challenged and my approach is that of an artist. The whole intent of this topic is as an art experiment. a continuation of the conversation, as art is meant to be. In this way, the soulless chimera or mimic

I've gone back and forth on whether to explain the patterns because it isn't exactly on the topic of the thread though it is relevant to your first post. Since other people seem confused I guess I'll go ahead, maybe it'll give you a springboard for your ideas...

Mimic is a somewhat accurate way to describe AI... it's just pattern recognition creating a pattern that's asked of it. For instance your original example images, apparently it sees the most prominent features of a dog pattern as its nose, and a parrot pattern as its eye. That's why it can't draw an archer's bow, because it doesn't know the concept of a string - to the AI it's just two lines that often appear in images associated with the word bow, and sometimes meet near fingers or are drawn over a face.

That's why your initial examples were a bit confusing as a topic starter because they aren't subversive or meta, it's just giving you what you asked for and you as the human are applying a meaning or relevance to it - but if you look closer it was just being very literal about your request. The weird pattern is the ripples you requested, and the hidden dogs, many with only their nose exposed. The second image is the same idea but it looks like hidden parrots with only the eye seen in most copies.

High level concepts like "hallucinating" or "algorithm" are largely ignored, printed as text (as seen in the negation example I showed), or turned into a synonym that it knows how to apply as a pattern (potentially it could be seen as a color palate, or a placement technique respectively, for instance). So while your mind may see the pattern as hallucination, it's an anthropomorphizing of the AI that's basically just a hash table that maps words to patterns derived from seeing thousands of the same image.

So it might have seen a horse you drew on deviantart, along with hundreds of thousands of other horse images, and knows the approximate shape and varying "mr potatohead" visual features like a mane that sometimes appear next to a shape it knows as the word horse. And it can make you that shape in all kinds of different styles, but at the end of the day it all boils down to a machine using a stamp with the word "horse" on it - except the stamp is a bit like a random character generator in a video game, with all the dials being tweaked on each pull of the lever.

So I think you kind of end up circling around and using AI to analyze humanity's patterns, as with @WilliamRedRobin's interesting analysis. Because any bias or intention you see in the AI "stamps" are a representation of a human pattern (or even of the mind of the interpreter exposing how they see the world). I guess in your case one of those would be the idea that art has a soul, rather than seeing AI as simply a question of copyright or not.

 

[tremault]

I've gone back and forth on whether to explain the patterns because it isn't exactly on the topic of the thread though it is relevant to your first post. Since other people seem confused I guess I'll go ahead, maybe it'll give you a springboard for your ideas...

Mimic is a somewhat accurate way to describe AI... it's just pattern recognition creating a pattern that's asked of it. For instance your original example images, apparently it sees the most prominent features of a dog pattern as its nose, and a parrot pattern as its eye. That's why it can't draw an archer's bow, because it doesn't know the concept of a string - to the AI it's just two lines that often appear in images associated with the word bow, and sometimes meet near fingers or are drawn over a face.

That's why your initial examples were a bit confusing as a topic starter because they aren't subversive or meta, it's just giving you what you asked for and you as the human are applying a meaning or relevance to it - but if you look closer it was just being very literal about your request. The weird pattern is the ripples you requested, and the hidden dogs, many with only their nose exposed. The second image is the same idea but it looks like hidden parrots with only the eye seen in most copies.

High level concepts like "hallucinating" or "algorithm" are largely ignored, printed as text (as seen in the negation example I showed), or turned into a synonym that it knows how to apply as a pattern (potentially it could be seen as a color palate, or a placement technique respectively, for instance). So while your mind may see the pattern as hallucination, it's an anthropomorphizing of the AI that's basically just a hash table that maps words to patterns derived from seeing thousands of the same image.

So it might have seen a horse you drew on deviantart, along with hundreds of thousands of other horse images, and knows the approximate shape and varying "mr potatohead" visual features like a mane that sometimes appear next to a shape it knows as the word horse. And it can make you that shape in all kinds of different styles, but at the end of the day it all boils down to a machine using a stamp with the word "horse" on it - except the stamp is a bit like a random character generator in a video game, with all the dials being tweaked on each pull of the lever.

So I think you kind of end up circling around and using AI to analyze humanity's patterns, as with @WilliamRedRobin's interesting analysis. Because any bias or intention you see in the AI "stamps" are a representation of a human pattern (or even of the mind of the interpreter exposing how they see the world). I guess in your case one of those would be the idea that art has a soul, rather than seeing AI as simply a question of copyright or not.

Thanks, this is the kind of exploration and communication I'm seeking. I appreciate your feedback on my initial post. I guess this could be oversight on my behalf. The prompt is very much tailored to produce commentary on various things I've experienced with AI and perhaps mistakenly believed were common knowledge. For example, the earliest ai generated images tended to resemble fever dreams or psychedelic trips with recurring patterns of dogs and birds. Also, during many of the discussions about AI like on technology podcasts, there has been much talk on how these ai algorithms can end up 'hallucinating' after some time and end up delivering garbage data.

edit : https://www.sciencealert.com/these-trippy-images-show-how-google-s-ai-sees-the-world-read-more

 

[unfair]

there has been much talk on how these ai algorithms can end up 'hallucinating' after some time and end up delivering garbage data.

The most recent instance of this was a new update and basically they just rolled back to an old model when the weird behavior started popping up. I don't know if it's how Microsoft does it, but one method of training these AI is basically natural selection, you iteratively pick the ones that learn to create the content in the way you want, and toss out the rest that are generating undesirable patterns. It's most easily seen in teaching an AI to walk. Essentially the AI is randomly trying thousands of different movements and the ones that are most successful get selected for as "learning".

So in that sense the hallucinating is basically a bug in the system akin to a memory overflow, it's jumping to another spot in its code to generate output in an unexpected way, perhaps even a hidden old iteration still secreted away inside the black box code that makes up the AI. That's the weird part, we don't quite understand the pattern definitions it uses, we select for the output not the code base. It's a million monkeys with typewriters and we picked the one that managed to write a word, and then a future iteration that wrote a sentence, and thousand iterations later it wrote a novel. But in the end it's just an efficient way of finding the one monkey that's the best at typing output we like without understanding the output.

 

[DigModiFicaTion]

I can't always explain exactly how I feel. We often explore how we feel through art. I'm not disproving anything. This isn't a science paper, It's communication of things that I can't readily put into words. Art is about exploring concepts and pushing against boundaries etc.
I'd argue btw, that not everyone here is necessarily an artist, there is a difference between art and design. There are surely lots of artists here and lots of designers but not everyone who is one is also the other. it's nebulous.

That kind of reads like art isn't art unless it's art, but you have to be an artist to understand and produce art. Seems exclusive. Perhaps art is simple in nature as the most natural form of expression and everyone can and is an artist without even knowing it. Those who focus on the mediums that are considered art are labeled artists, but I don't believe art is exclusive to that group. Inaccessible art is simply nebulous expression. I've seen many examples of what people call art that don't fit my definition, but it's still their art.

 

[Possessed]

That kind of reads like art isn't art unless it's art

And as we all know, art is dead.

 

[tremault]

That kind of reads like art isn't art unless it's art, but you have to be an artist to understand and produce art. Seems exclusive. Perhaps art is simple in nature as the most natural form of expression and everyone can and is an artist without even knowing it. Those who focus on the mediums that are considered art are labeled artists, but I don't believe art is exclusive to that group. Inaccessible art is simply nebulous expression. I've seen many examples of what people call art that don't fit my definition, but it's still their art.

Anyone can do art, but not everything everyone does is art. I suppose you could look at it as a process. If a graphic designer uses the graphic design process to produce a logo, is that art?no, it's graphic design, specifically logo design. It's a craft for sure. I'm just using that description of a method of working that expresses thoughts and emotions in a compelling and thoughtful way. Someone throwing paint at a canvas and thinking it looks like a funny pattern isn't art and it doesn't end up in galleries, but a person doing a lot of research and going through a process to get in a head space and throw a specific type of pain, in a specific way, in a particular setting, in order to tell a story and communicate a message, can be considered art, if people are understanding and getting some insight through it.
I mean if I just do a random image generation now, and type in a clown with a ball, that's taken me 2 seconds to think about and I'm communicating nothing. Even if the clown looks odd or interesting or funny, it doesn't matter because it's not communicating any thoughts, research or feelings. It has no meaning.

This image is garbage. I put less thought into it than a crappy doodle I did while in Spanish class yesterday. It looks visually accomplished but it has no soul at all. It's as valuable as what comes out of my nose.

 

[Gibichung]

Another AI thread. This has been my stance since the beginning.

As for messing around or even use it for fanedits? That's a different thing.

I totally encourage messing around just like @tremault is doing here and everyone else. I'd treat it the same way as fan-editing. As I mentioned on the other thread, I think it's cool to use as fanedit cover-art, I think the least anyone can do is find a way to credit the original artists and/or prompts somewhere (like even as small as a caption, maybe even separate from the AI generated piece itself) for the images used. It's the same way I've seen many folks approach their posters. I've seen incredible fan-art done by artists who aren't the fan-editor, and I think that's so cool when artists are credited. I think same thing applies. The ground rules here is that we don't sell or make any money from it, so as far as I'm considered it's fanart... just like fanediting.

 

[Gibichung]

And as we all know, art is dead.

I wish I can like this 100x lol

topical!!!

 

[WilliamRedRobin]

I would argue that value has more to do with what you get out of something than what you put into it. And if something is valuable to one person, then it's valuable.
As for what is or isn't art, I don't really care. It's a word with no set meaning, which makes it meaningless. I want nothing to do with art. I just want to make things.
And AI is a great tool for making things.

 

[Moe_Syzlak]

I haven’t read all of this but I’d say AI can certainly be art and seems to me to be another tool. How that tool is used is up to the artist. It seems somewhat analogous discussions about whether sampling constitutes original music. I think there’s some music on something like Paul’s Boutique that uses sampling very much to create unique art. But then there’s songs like Puff Daddy’s version of Every Breath You Take which is no more than a cover song to me, yet he calls it an original (albeit with hefty royalties to Sting). Every time there are new tools there’s always pushback and always abuse. Usually things sort themselves out organically.