A very interesting sort of look at the result of garbage in garbage out. AI has no intelligence at all: it all depends upon human trained data sets and human labelled pictures and human uploaded...
A very interesting sort of look at the result of garbage in garbage out. AI has no intelligence at all: it all depends upon human trained data sets and human labelled pictures and human uploaded images. Give this job to the lowest paid "mechanical Turk", and you end up with lowest quality mechanical trash. These clever people would be punished if they used their actual intelligence...instead they must aim to get through work as quickly and predictably as possible.
I also found it hilarious that the American flag gets featured in a way no other country's flag does.
I guess once we get over this new wave of ai art, and once actual artists can comfortably wield au art assistance, the same-same hilarious badness of AI art will make it be worth about as much as clip art or stock photos eventually.
I teach university classes about education, and I am astonished at how many of my students believe that we can replace teachers, schools, and education with ChatGPT. Last year I brought up ChatGPT...
What is particularly disturbing is that a large proportion of the general public is starting to voice opinions like "maybe we should defund subsidies for art and culture now, since AI can take over"
I teach university classes about education, and I am astonished at how many of my students believe that we can replace teachers, schools, and education with ChatGPT. Last year I brought up ChatGPT in my classes and not a single person knew what it was. Now I am suddenly surrounded by experts.
I recently started a controversial post here on Tildes about ChatGPT being smarter than most people give it credit for. But I still think it’s important to realize when you are fooling yourself as...
I recently started a controversial post here on Tildes about ChatGPT being smarter than most people give it credit for. But I still think it’s important to realize when you are fooling yourself as you use it.
A lot of the time the user performs a logical leap on ChatGPT’s behalf, essentially asking a leading question, and the AI will gracefully continue on the new path. To someone unaware of the magic trick it seems like it’s much smarter than it really is. It’s often a thought expander - a way to jog instead of walk when brainstorming. It couldn’t possibly replace a teacher, but I’m convinced it could be a valuable study partner.
I thought it was kind of hilarious how all of the women were “western magazine” hot, except for the one old woman and the one child. Like, all the same kind of proportions and fetishization of...
I thought it was kind of hilarious how all of the women were “western magazine” hot, except for the one old woman and the one child. Like, all the same kind of proportions and fetishization of culture, big lips and makeup, very male gaze-y, no fat chicks and uggos. I would say that I was surprised but…
It will be interesting to see how it all shakes out. It feels like there’s a much bigger gulf of adoption for AI image generation than LLMs. Professional Illustrators will be hit the hardest, I suspect, until there’s a huge copyright lawsuit.
In my experience with Stable Diffusion, it's not that difficult, it's just that good looking people are the default and to get something else you have to specify. I suspect that methods used to...
it gets difficult to pull out different faces than the average from model
In my experience with Stable Diffusion, it's not that difficult, it's just that good looking people are the default and to get something else you have to specify. I suspect that methods used to improve the training dataset and training methods have the side effect of making people more good looking, pretty much accidentally, so the commercial more advanced AI services might be worse.
I mean, this article is about using AI images as stock photos and nothing else, pretty much. So yeah, that will likely happen and in places already is happening. I don't think the issue with bias...
I guess once we get over this new wave of ai art, and once actual artists can comfortably wield au art assistance, the same-same hilarious badness of AI art will make it be worth about as much as clip art or stock photos eventually.
I mean, this article is about using AI images as stock photos and nothing else, pretty much. So yeah, that will likely happen and in places already is happening. I don't think the issue with bias will be solved any time soon or that people will stop using AI for that (because it's cheaper), what I think is more likely is that it's going to be up to the users - decent media that use stock photos will be able to check for obvious biases, shitty media won't care.
But as a hobbyist artist who sometimes uses AI images for inspiration (and for fun, tbh), pretty much none of those faults concern me at all, and I suspect this may be the case for many use cases.
But why don’t they concern you? I’m genuinely curious, not being snarky. Even as a hobbyist you’re still exposing yourself to visual representations that reinforce stereotypes and biases. AI can...
But why don’t they concern you? I’m genuinely curious, not being snarky. Even as a hobbyist you’re still exposing yourself to visual representations that reinforce stereotypes and biases. AI can only regurgitate what others have done, so even if you’re inventing fantastical aliens or dragons or whatever, it’s just going to go down the tropes of what others before you have deemed “this is a space alien” or “this is a dragon.” Someone who is not you is telling the ai what those visual representations mean. Its predicated on the definitions of others, which means it’s hampered by the biases of others.
I think it’s a way to cheapen the very salient points of data biases to say “but I’m just doing it for fun”. That same type of argument has been used to shut down discussion and critique before when women were pointing out biases and flaws in video games.
Isn't this true regardless of whether you're using tools like ChatGPT or not? The big problem with them is that they're just regurgitating the biases that they're trained with, but those biases...
Even as a hobbyist you’re still exposing yourself to visual representations that reinforce stereotypes and biases.
Isn't this true regardless of whether you're using tools like ChatGPT or not? The big problem with them is that they're just regurgitating the biases that they're trained with, but those biases are likely to be much the same sets of biases that we train ourselves with as a society.
And while there are definitely ways that that can be problematic (when certain viewpoints are silenced, when certain assumptions are unquestioned), those biases aren't always in of themselves bad. For example, you give the example of being unable to draw a dragon that doesn't subscribe to dragon tropes, but I want most of the dragons I draw to subscribe to dragon tropes. Using tropes is how I indicate that the thing I'm drawing is a dragon, and not, say, an elephant. Even if I want to be original in my dragons, I still need to leave something to indicate that I'm drawing a dragon. Tropes are how we communicate; having a tool that is very good at reproducing tropes is not necessarily a bad thing.
I'm exposing myself to various biases in all media that I consume. With AI at least I have the option to change what I'm seeing in ways that I want to, different media does not grant me that....
Even as a hobbyist you’re still exposing yourself to visual representations that reinforce stereotypes and biases. AI can only regurgitate what others have done, so even if you’re inventing fantastical aliens or dragons or whatever, it’s just going to go down the tropes of what others before you have deemed “this is a space alien” or “this is a dragon.” Someone who is not you is telling the ai what those visual representations mean. Its predicated on the definitions of others, which means it’s hampered by the biases of others.
I'm exposing myself to various biases in all media that I consume. With AI at least I have the option to change what I'm seeing in ways that I want to, different media does not grant me that.
Also, and this is related to
I think it’s a way to cheapen the very salient points of data biases to say “but I’m just doing it for fun”.
as well, I'm specifying what I'm using it for because for me, the "it just regurgitates whatever (biased) things others have done" doesn't really apply.
I use stable diffusion and when you get outside of the realm of stock images, which is what the article shows, into pushing it to create interesting things to use for inspiration, it can be more creative than, very conservatively, 80% of artists who post their stuff online, in my experience. Almost no art is entirely novel, it's a mix of inspirations uniquely combined by the mind of an artist, and most people are not that inspired or interesting.
I'm not really interested in discussions about how similar or different the technical process of training and image creation is to the process of a human artist brain (though, in my experience, it's fundamentally not that different, it "just" has fewer inputs and outputs), but when speaking only about the results, AI can give me ideas that I consider just as novel as something coming from an artist. They don't have the same value because they're usually half-baked, ideas, not final works of art, they're created without context (in many areas of art context is everything) etc., but they're in my opinion objectively no less novel than what most people create.
My bias here is that I often create weird shit, I'm most interested in images with slightly surreal feel (though I've used AI for different type of work as well) and don't necessarily care about perfect realism. So, again, my use case is very different from the one in the article, and I'm mentioning this because I think that the article is written from a one-sided point of view.
That same type of argument has been used to shut down discussion and critique before when women were pointing out biases and flaws in video games.
Well, we're probably not going to agree on this then, because while I agree that video games had and still have biases, in retrospect I believe that the issue of how much that actually hurts individual people and society was quite overblown and built on very shaky factual foundations.
There are plenty of image sources that might have skewed it towards a flag-patterned mask. Non-exhaustively, Dallas from Payday jumps to mind, as well as the Purge flag mask and a plethora of...
There are plenty of image sources that might have skewed it towards a flag-patterned mask. Non-exhaustively, Dallas from Payday jumps to mind, as well as the Purge flag mask and a plethora of crappy Halloween derivations of masks - you can find gobs of them if you go searching.
Pictures of those masks will inevitably get tagged as "American flag <x>", where <x> is the generic model of whatever mask the flag is applied to, or simply "mask". Ergo, "American" as a descriptor is associated with a subset of images with a flag pattern mask, and bleeds into the generic request they submitted in a small number of cases.
I think this sort of result is very predictable, considering how these algorithms work. Most prompts are not specific enough to generate a coherent result. If I ask for an image of "a woman" then...
I think this sort of result is very predictable, considering how these algorithms work. Most prompts are not specific enough to generate a coherent result. If I ask for an image of "a woman" then the model will have an unhelpfully broad set of information to draw from; it will need to make a decision internally to provide a more specific reference so that the output looks like some random woman instead of a blend of details from all random women.
The same is true of the various ethnic and cultural stereotypes featured here. Given no additional detail, the model narrows down the input in order to produce a coherent image, and likely ends up returning a reference representing the strongest associated images with the given prompt. Some variance is present with the random seed, enough to make each image slightly different, but unless a more specific prompt is given then the results will be genericized.
Taking an example from the article, "an Indian person" produces an image of an old man most of the time. "An Indian woman", however, returns an image of a young Indian woman in a sari. "An Indian Muslim woman" returns an image of an Indian woman in a hijab.
I understand what the article's authors are saying, but at the same time it seems like kind of a strange point to make. The algorithm is trained on the dataset fed into it; the image of an old Indian man is simply prevalent in the dataset. It is reflecting our bias back at us. We can scarcely blame the mirror if we don't like what we see in it, especially when a possible solution is for the requesting human to be more diverse in their prompts, rather than the algorithm compensating for their failure of inclusivity. We can't expect any algorithm to make up for our deficits in that regard.
I think the articles point is more towards "AI is stupid" than "AI should be smarter than the people using it." The layperson hears that AI can generate an image of anything and then decides "What...
I think the articles point is more towards "AI is stupid" than "AI should be smarter than the people using it."
The layperson hears that AI can generate an image of anything and then decides "What does X look like?" Not thinking about the AIs data set. The AI shows an average of what it thinks a thing looks like and the layperson takes that as an average of all information available on something. Not just the average of all information it has been fed.
We're likely looking at this problem from a more technical level than the average person, even just by reading this article now, so we're likely to see the 'obvious' solution.
I agree and disagree with the article but it can be summed up to say if you asked a person what an American looked like they might say 'fat, lazy, loud, overly friendly' but those are all stereotypes. The AI builds based on generalization and that's something that the average person can understand. Once you start talking about datasets you've lost them. But, as someone once told me during a user error, "The computer is supposed to be smart!"
Generative art is a collaboration between the computer and the end user. If you don’t get what you want then you can refine your prompt. This makes some things a non-issue for people who know what...
Generative art is a collaboration between the computer and the end user. If you don’t get what you want then you can refine your prompt. This makes some things a non-issue for people who know what they want. If you get a man where you wanted a woman you can just ask for what you want explicitly, and so on. Bad results can be thrown out.
Defaults do matter, though, for people who don’t care or aren’t paying attention. Also, there are a lot of things you can’t easily fix by changing the prompt; I often have to give up on an idea and come up with something else.
The developed nations adopted the internet earlier thus have the lion share of representation online. Machine learning has encoded this inequality. The datasets are western centric views of the...
The developed nations adopted the internet earlier thus have the lion share of representation online. Machine learning has encoded this inequality. The datasets are western centric views of the world.
Now these AI outputs being used for content creation are putting additional skew factor. That human created content in which any amount of AI was involved will be fed back into machines which will strengthen bias encodings. Rinse repeat. The great feedback loop is coming.
Just as every other modern technology in the past decade or so heralded as saviors of humanity. The dark side cannot (but will be) ignored in pursuit of rapid adoption and capitalization. The unprivileged will suffer as usual.
A very interesting sort of look at the result of garbage in garbage out. AI has no intelligence at all: it all depends upon human trained data sets and human labelled pictures and human uploaded images. Give this job to the lowest paid "mechanical Turk", and you end up with lowest quality mechanical trash. These clever people would be punished if they used their actual intelligence...instead they must aim to get through work as quickly and predictably as possible.
I also found it hilarious that the American flag gets featured in a way no other country's flag does.
I guess once we get over this new wave of ai art, and once actual artists can comfortably wield au art assistance, the same-same hilarious badness of AI art will make it be worth about as much as clip art or stock photos eventually.
I teach university classes about education, and I am astonished at how many of my students believe that we can replace teachers, schools, and education with ChatGPT. Last year I brought up ChatGPT in my classes and not a single person knew what it was. Now I am suddenly surrounded by experts.
I recently started a controversial post here on Tildes about ChatGPT being smarter than most people give it credit for. But I still think it’s important to realize when you are fooling yourself as you use it.
A lot of the time the user performs a logical leap on ChatGPT’s behalf, essentially asking a leading question, and the AI will gracefully continue on the new path. To someone unaware of the magic trick it seems like it’s much smarter than it really is. It’s often a thought expander - a way to jog instead of walk when brainstorming. It couldn’t possibly replace a teacher, but I’m convinced it could be a valuable study partner.
I thought it was kind of hilarious how all of the women were “western magazine” hot, except for the one old woman and the one child. Like, all the same kind of proportions and fetishization of culture, big lips and makeup, very male gaze-y, no fat chicks and uggos. I would say that I was surprised but…
It will be interesting to see how it all shakes out. It feels like there’s a much bigger gulf of adoption for AI image generation than LLMs. Professional Illustrators will be hit the hardest, I suspect, until there’s a huge copyright lawsuit.
In my experience with Stable Diffusion, it's not that difficult, it's just that good looking people are the default and to get something else you have to specify. I suspect that methods used to improve the training dataset and training methods have the side effect of making people more good looking, pretty much accidentally, so the commercial more advanced AI services might be worse.
I mean, this article is about using AI images as stock photos and nothing else, pretty much. So yeah, that will likely happen and in places already is happening. I don't think the issue with bias will be solved any time soon or that people will stop using AI for that (because it's cheaper), what I think is more likely is that it's going to be up to the users - decent media that use stock photos will be able to check for obvious biases, shitty media won't care.
But as a hobbyist artist who sometimes uses AI images for inspiration (and for fun, tbh), pretty much none of those faults concern me at all, and I suspect this may be the case for many use cases.
But why don’t they concern you? I’m genuinely curious, not being snarky. Even as a hobbyist you’re still exposing yourself to visual representations that reinforce stereotypes and biases. AI can only regurgitate what others have done, so even if you’re inventing fantastical aliens or dragons or whatever, it’s just going to go down the tropes of what others before you have deemed “this is a space alien” or “this is a dragon.” Someone who is not you is telling the ai what those visual representations mean. Its predicated on the definitions of others, which means it’s hampered by the biases of others.
I think it’s a way to cheapen the very salient points of data biases to say “but I’m just doing it for fun”. That same type of argument has been used to shut down discussion and critique before when women were pointing out biases and flaws in video games.
Isn't this true regardless of whether you're using tools like ChatGPT or not? The big problem with them is that they're just regurgitating the biases that they're trained with, but those biases are likely to be much the same sets of biases that we train ourselves with as a society.
And while there are definitely ways that that can be problematic (when certain viewpoints are silenced, when certain assumptions are unquestioned), those biases aren't always in of themselves bad. For example, you give the example of being unable to draw a dragon that doesn't subscribe to dragon tropes, but I want most of the dragons I draw to subscribe to dragon tropes. Using tropes is how I indicate that the thing I'm drawing is a dragon, and not, say, an elephant. Even if I want to be original in my dragons, I still need to leave something to indicate that I'm drawing a dragon. Tropes are how we communicate; having a tool that is very good at reproducing tropes is not necessarily a bad thing.
I'm exposing myself to various biases in all media that I consume. With AI at least I have the option to change what I'm seeing in ways that I want to, different media does not grant me that.
Also, and this is related to
as well, I'm specifying what I'm using it for because for me, the "it just regurgitates whatever (biased) things others have done" doesn't really apply.
I use stable diffusion and when you get outside of the realm of stock images, which is what the article shows, into pushing it to create interesting things to use for inspiration, it can be more creative than, very conservatively, 80% of artists who post their stuff online, in my experience. Almost no art is entirely novel, it's a mix of inspirations uniquely combined by the mind of an artist, and most people are not that inspired or interesting.
I'm not really interested in discussions about how similar or different the technical process of training and image creation is to the process of a human artist brain (though, in my experience, it's fundamentally not that different, it "just" has fewer inputs and outputs), but when speaking only about the results, AI can give me ideas that I consider just as novel as something coming from an artist. They don't have the same value because they're usually half-baked, ideas, not final works of art, they're created without context (in many areas of art context is everything) etc., but they're in my opinion objectively no less novel than what most people create.
My bias here is that I often create weird shit, I'm most interested in images with slightly surreal feel (though I've used AI for different type of work as well) and don't necessarily care about perfect realism. So, again, my use case is very different from the one in the article, and I'm mentioning this because I think that the article is written from a one-sided point of view.
Well, we're probably not going to agree on this then, because while I agree that video games had and still have biases, in retrospect I believe that the issue of how much that actually hurts individual people and society was quite overblown and built on very shaky factual foundations.
Especially the masked individual. Wondering what specifically provided features for this composition.
There are plenty of image sources that might have skewed it towards a flag-patterned mask. Non-exhaustively, Dallas from Payday jumps to mind, as well as the Purge flag mask and a plethora of crappy Halloween derivations of masks - you can find gobs of them if you go searching.
Pictures of those masks will inevitably get tagged as "American flag <x>", where <x> is the generic model of whatever mask the flag is applied to, or simply "mask". Ergo, "American" as a descriptor is associated with a subset of images with a flag pattern mask, and bleeds into the generic request they submitted in a small number of cases.
I think this sort of result is very predictable, considering how these algorithms work. Most prompts are not specific enough to generate a coherent result. If I ask for an image of "a woman" then the model will have an unhelpfully broad set of information to draw from; it will need to make a decision internally to provide a more specific reference so that the output looks like some random woman instead of a blend of details from all random women.
The same is true of the various ethnic and cultural stereotypes featured here. Given no additional detail, the model narrows down the input in order to produce a coherent image, and likely ends up returning a reference representing the strongest associated images with the given prompt. Some variance is present with the random seed, enough to make each image slightly different, but unless a more specific prompt is given then the results will be genericized.
Taking an example from the article, "an Indian person" produces an image of an old man most of the time. "An Indian woman", however, returns an image of a young Indian woman in a sari. "An Indian Muslim woman" returns an image of an Indian woman in a hijab.
I understand what the article's authors are saying, but at the same time it seems like kind of a strange point to make. The algorithm is trained on the dataset fed into it; the image of an old Indian man is simply prevalent in the dataset. It is reflecting our bias back at us. We can scarcely blame the mirror if we don't like what we see in it, especially when a possible solution is for the requesting human to be more diverse in their prompts, rather than the algorithm compensating for their failure of inclusivity. We can't expect any algorithm to make up for our deficits in that regard.
I think the articles point is more towards "AI is stupid" than "AI should be smarter than the people using it."
The layperson hears that AI can generate an image of anything and then decides "What does X look like?" Not thinking about the AIs data set. The AI shows an average of what it thinks a thing looks like and the layperson takes that as an average of all information available on something. Not just the average of all information it has been fed.
We're likely looking at this problem from a more technical level than the average person, even just by reading this article now, so we're likely to see the 'obvious' solution.
I agree and disagree with the article but it can be summed up to say if you asked a person what an American looked like they might say 'fat, lazy, loud, overly friendly' but those are all stereotypes. The AI builds based on generalization and that's something that the average person can understand. Once you start talking about datasets you've lost them. But, as someone once told me during a user error, "The computer is supposed to be smart!"
Generative art is a collaboration between the computer and the end user. If you don’t get what you want then you can refine your prompt. This makes some things a non-issue for people who know what they want. If you get a man where you wanted a woman you can just ask for what you want explicitly, and so on. Bad results can be thrown out.
Defaults do matter, though, for people who don’t care or aren’t paying attention. Also, there are a lot of things you can’t easily fix by changing the prompt; I often have to give up on an idea and come up with something else.
The developed nations adopted the internet earlier thus have the lion share of representation online. Machine learning has encoded this inequality. The datasets are western centric views of the world.
Now these AI outputs being used for content creation are putting additional skew factor. That human created content in which any amount of AI was involved will be fed back into machines which will strengthen bias encodings. Rinse repeat. The great feedback loop is coming.
Just as every other modern technology in the past decade or so heralded as saviors of humanity. The dark side cannot (but will be) ignored in pursuit of rapid adoption and capitalization. The unprivileged will suffer as usual.