You can prompt an image genrater to just spit out the original art it trained on.
Imagine I had been classically trained as a painter. I study works from various artists. I become so familiar with those works - and skilled as a renderer of art in my own right - that I can reproduce, say, the Mona Lisa from memory with exacting accuracy. Should I be allowed to claim it as my art? Sign my name to it? Sell it as my own?
Now lets say we compare the original and my work at the micron level. I’m human, there’s no way I can match the original stroke for stroke, bristle to bristle. However small, there are differences. When does the work become transformative?
Let’s switch to an image generator. I ask for a picture of a smiling woman, renaissance style. The model happens to be biased to DaVinci, and it spits out almost exactly the same work as the Mona Lisa. Let’s say as a prompt engineer, I’ve never heard of or seen the Mona Lisa. I take the image, decide “meh, good enough for what I need right now”, and use it in some commercial product (say, a t-shirt). Should I be able to do that? What if it’s not the Mona Lisa, it’s a work from a living artist?
What if it’s not an image? Say I tell some model to make a song and it accidentally produces Greenday’s Basketcase (which itself is basically just a modified Pachelbel’s Canon), can I put that on a record and sell it? Who’s responsibility is it to make sure that a model’s output is unique or transformative? Shit, look at all the legal cases where musicians are suing other musicians because the chord progression is similar in two songs; What happens when it’s exactly the same because the prompt engineer for a music generation model isn’t paying attention?
You might have noticed that I haven’t referred to this technology as AI. That’s because it’s not. It’s Machine Learning. It has no intelligence. It neither seeks to create beautiful, original art, nor does it intend to rip someone off. It has no plans, no aspirations, no context, no whims. It’s a parrot, spitting out copies of things we ask it for. In general, these outputs are mixtures of various things, but sometimes they aren’t. They just output some of the training data, because that’s the output that - statistically - was the best match for the prompt.
As an artist myself, I don’t fear machine learned models. I fear that these greedy fuckin’ companies will warehouse any and every bit of data they can get their hands on, train their models on other people’s work, never pay them a dime, and rip off the essence of their art without any regard for what will happen to the original artists after some jackass execs tell all their advertising/webdesign/programming/scriptwriting/etc departments to just ask the “AI” to “design” everything.
You can already see this happening with game studios. Writers went on strike over it.
You can prompt an image genrater to just spit out the original art it trained on.
This is a common misconception. It’s not true, except in the extremely rare case of “overfitting” that all AI trainers work very hard to avoid. It’s considered a bug, because why would anyone spend millions of dollars and vast computer resources poorly replicating what can be accomplished with a simple copy/paste operation? That completely misses the point of all this.
If an AI art generator spits out copies of its training data it’s a failure of an AI art generator.
You can prompt an image genrater to just spit out the original art it trained on.
This is incorrect. Have you tried doing it?
That’s not how AI works. It’s not magic, nor does it create “copies”. It creates entirely original works, with influences from other works (similar to what other humans do).
@burliman
You can prompt an image genrater to just spit out the original art it trained on.
Imagine I had been classically trained as a painter. I study works from various artists. I become so familiar with those works - and skilled as a renderer of art in my own right - that I can reproduce, say, the Mona Lisa from memory with exacting accuracy. Should I be allowed to claim it as my art? Sign my name to it? Sell it as my own?
Now lets say we compare the original and my work at the micron level. I’m human, there’s no way I can match the original stroke for stroke, bristle to bristle. However small, there are differences. When does the work become transformative?
Let’s switch to an image generator. I ask for a picture of a smiling woman, renaissance style. The model happens to be biased to DaVinci, and it spits out almost exactly the same work as the Mona Lisa. Let’s say as a prompt engineer, I’ve never heard of or seen the Mona Lisa. I take the image, decide “meh, good enough for what I need right now”, and use it in some commercial product (say, a t-shirt). Should I be able to do that? What if it’s not the Mona Lisa, it’s a work from a living artist?
What if it’s not an image? Say I tell some model to make a song and it accidentally produces Greenday’s Basketcase (which itself is basically just a modified Pachelbel’s Canon), can I put that on a record and sell it? Who’s responsibility is it to make sure that a model’s output is unique or transformative? Shit, look at all the legal cases where musicians are suing other musicians because the chord progression is similar in two songs; What happens when it’s exactly the same because the prompt engineer for a music generation model isn’t paying attention?
You might have noticed that I haven’t referred to this technology as AI. That’s because it’s not. It’s Machine Learning. It has no intelligence. It neither seeks to create beautiful, original art, nor does it intend to rip someone off. It has no plans, no aspirations, no context, no whims. It’s a parrot, spitting out copies of things we ask it for. In general, these outputs are mixtures of various things, but sometimes they aren’t. They just output some of the training data, because that’s the output that - statistically - was the best match for the prompt.
As an artist myself, I don’t fear machine learned models. I fear that these greedy fuckin’ companies will warehouse any and every bit of data they can get their hands on, train their models on other people’s work, never pay them a dime, and rip off the essence of their art without any regard for what will happen to the original artists after some jackass execs tell all their advertising/webdesign/programming/scriptwriting/etc departments to just ask the “AI” to “design” everything.
You can already see this happening with game studios. Writers went on strike over it.
This is a common misconception. It’s not true, except in the extremely rare case of “overfitting” that all AI trainers work very hard to avoid. It’s considered a bug, because why would anyone spend millions of dollars and vast computer resources poorly replicating what can be accomplished with a simple copy/paste operation? That completely misses the point of all this.
If an AI art generator spits out copies of its training data it’s a failure of an AI art generator.
This is incorrect. Have you tried doing it?
That’s not how AI works. It’s not magic, nor does it create “copies”. It creates entirely original works, with influences from other works (similar to what other humans do).