Except when it's a literal copy of something that previously existed.
It isn't a copy. That isn't how that works. Deep Learning Neural Networks don't contain a library of images.
They are trained with a library of images, which is then taken away.
Convolutional Neural Networks play the Hot and Cold game for head-pats.
The training environment gives the Neural Network a prompt. For example "Draw a squid."
The NN does its best to draw a squid. The result is compared to one of the training images from the library of squid images. The NN is awarded head-pats accordingly.
This repeats thousands or millions of times.
A well trained AI can do things like drawing a well known character as though they were a squid which will combine their ability to satisfy both sets of criteria at the same time.
Like it or not, that image has never been created before.
That's not a serious description and demonstrates a poor grasp on the underlying technology. It's guided recursive perturbation of noise. Until that noise looks enough like something it's seen before to become classified as the same.
Anyone else would call this copying with a little bit of randomness added in. You should really dig a little deeper here.
Anyways try the exclusive case. Have it draw something that has never been classified before. You can look around for items of antiquity or lost cultures and see how it fares. Since it can't do these things it obviously only can do things it has seen and have been classified. So it can only reproduce what it has been trained on. This is obvious.
which will combine their ability to satisfy both sets of criteria at the same time.
Wow, so it combined two copies into one thing? That's still copying.
It is vaguely impressive that preturbative systems are effective at quickly finding approximate solutions. Which is why "satisfying both sets of criteria" means you get 7 fingered hands and other oddities expressed in the image. Add more criteria for more fun.
You're really just playing Monte Carlo on a stolen deck of classifications. To be fair, Google did pay a bunch of africans $2/day to tag the images, so there is some originality in the database, but very little.
that image has never been created before.
Fortunately that's now how copyright works or how copying is defined in the eyes of the law.
Anyone else would call this copying with a little bit of randomness added in.
Yeah, so we are having an issue of semantics here. We are using different meanings of the word copy. Further, I think you know what the actual definition is, and you are arguing in bad faith.
If a human artist looks at a drawing or an image, then makes a copy of it by hand, it is substantially different. This can be sold as original art under a couple conditions. I won't get into the conditions here, because I don't think they matter and you don't care.
If the artwork is different enough that it won't be confused for commercial purposes, it doesn't infringe on copyright.
Moreover, there are things that can not be covered by copyright. These include styles and formats. There has been a Chinese version of the Australian show "Who Wants to Be a Millionaire." Which in turn is a copy of the CBS show "The $64,000 Question" from 1955, because you can't copyright show formats.
You also can't copyright procedures, rules or recipes. You also can't copyright an artistic school or style.
So, you are conflating the term "duplicate" with the term "copy". Duplicates are covered by copyright, as they are an exact reproduction of an artistic work.
A rabbi crawling out of a drain grate in the style of Studio Ghibli absolutely is not.
To take a photographic image and then redraw that into a new picture in an entirely different style is not a copy. It certainly isn't a duplicate.
To combine two images into a fusion of artistic representations into an image that has never existed before is not a copy.
Have it draw something that has never been classified before.
You are correct. A convolutional neural network will not be able to generate images that match criteria outside its training data set.
But you can't speak Chinese without every having learned it. So what? What does that have to do with anything at all?
Your major complaint seems to be that Convolutional Neural Networks are not, in fact, people. You got me! They have less neurons than a flatworm. They are certainly less complex than the nervous system of a garden slug. They get things wrong and draw stupid shit all the time.
That doesn't mean they are a Xerox machine. They have a different function and are really good at certain tasks, especially of classification. They are very scalable. They are not thinking machines in any sense of the word "thinking".
Further, I think you know what the actual definition is, and you are arguing in bad faith.
Really?
If a human artist looks at a drawing or an image, then makes a copy of it by hand, it is substantially different.
Transformative is the standard. And it can be adjudicated. Pure copies by hand are called forgeries and are not legal to sell. I don't know why you think a hand copy is immediately substantially transformative when there's centuries of case law saying precisely the opposite and offering real tests to be used in making the determination.
You also can't copyright an artistic school or style.
If I can recognize the style then it's obvious you copied it.
So, you are conflating the term "duplicate" with the term "copy".
No, you are, in an effort to argue in bad faith. The tool makes copies. Sometimes those copies are so bad they actually do violate copyright. That they can be transformative does not relieve them of being copies. This is obvious.
Your previous example seemed to rely on the fact you could combine a known character with octopus parts and call that a "new work." The point being, even copyright law, with all it's exceptions, would call this a violation. You can't draw Iron Man with squid legs and call it a new character.
That doesn't mean they are a Xerox machine.
Effectively they are. So are humans. The difference is humans are additionally capable of creating entirely new things. These toys cannot. They will not. There is no upgrade to make them capable of it. There is no obvious technological path from the current implementation to a new one which is equally capable.
This thing can only copy. To get back on point your 30 year timeline is bogus. If you understand this technology so well then you should see this clearly.
Guy, when you enter a prompt into a Convolutional Neural Network, you can't point to the work that it is coping; because it isn't copying a single work, or even a couple. The NN is relying on the complex interactions of neurons to identify patterns within its output and to adjust those to make them better match a gestalt of that training.
You know this. You tried to correct my original response to point this out to me.
When you duplicate an image with a Xerox machine, or even a person visually copying an artwork by looking at it, you can point to the piece being copied. "There it is, that is the original."
Duplication is prohibited under copywrite law; but style isn't. Drawing Spongebob Squarepants in the style of Vincent Van Gough is a creative, transformational work even if you can recognize both the character and the style.
There may be issues about selling or distributing that work, because Spongebob is covered by trademark etc., but there is no question about the transformative and artistic nature of the piece. Juxtaposing the vapid childrens' character and one of the most famous Dutch painters is, in itself, an artistic statement.
No, you cannot copyright an art style. While you can copyright specific, original works of art, such as paintings, drawings, or sculptures, copyright law does not protect styles or general ideas.
I get that you fucking hate Neural Network generated art. I understand, but you are just making shit up to win arguments. You are demonstrably wrong, which is worse because you know you are wrong, and you don't give a fuck.
Every day we make fun of Neo-Marxist fuckheads for making up their own facts, and here you are doing it to win an internet slap fight.
Even if you manage to declare victory by utterly distorting terms (a leftist tactic) you will have convinced no one of your point of view.
There are really, really good reasons to distrust Convolutional Neural Networks and Large Language Models. I've given a few very good ones in other posts in this topic. "lol; they are just Xerox machines!' isn't one.
Oh, and BTW, I've never given a 30 year prediction. That was someone else.
It isn't a copy. That isn't how that works. Deep Learning Neural Networks don't contain a library of images.
They are trained with a library of images, which is then taken away.
Convolutional Neural Networks play the Hot and Cold game for head-pats.
The training environment gives the Neural Network a prompt. For example "Draw a squid."
The NN does its best to draw a squid. The result is compared to one of the training images from the library of squid images. The NN is awarded head-pats accordingly.
This repeats thousands or millions of times.
A well trained AI can do things like drawing a well known character as though they were a squid which will combine their ability to satisfy both sets of criteria at the same time.
Like it or not, that image has never been created before.
That's not a serious description and demonstrates a poor grasp on the underlying technology. It's guided recursive perturbation of noise. Until that noise looks enough like something it's seen before to become classified as the same.
Anyone else would call this copying with a little bit of randomness added in. You should really dig a little deeper here.
Anyways try the exclusive case. Have it draw something that has never been classified before. You can look around for items of antiquity or lost cultures and see how it fares. Since it can't do these things it obviously only can do things it has seen and have been classified. So it can only reproduce what it has been trained on. This is obvious.
Wow, so it combined two copies into one thing? That's still copying.
It is vaguely impressive that preturbative systems are effective at quickly finding approximate solutions. Which is why "satisfying both sets of criteria" means you get 7 fingered hands and other oddities expressed in the image. Add more criteria for more fun.
You're really just playing Monte Carlo on a stolen deck of classifications. To be fair, Google did pay a bunch of africans $2/day to tag the images, so there is some originality in the database, but very little.
Fortunately that's now how copyright works or how copying is defined in the eyes of the law.
Yeah, so we are having an issue of semantics here. We are using different meanings of the word copy. Further, I think you know what the actual definition is, and you are arguing in bad faith.
If a human artist looks at a drawing or an image, then makes a copy of it by hand, it is substantially different. This can be sold as original art under a couple conditions. I won't get into the conditions here, because I don't think they matter and you don't care.
If the artwork is different enough that it won't be confused for commercial purposes, it doesn't infringe on copyright.
Moreover, there are things that can not be covered by copyright. These include styles and formats. There has been a Chinese version of the Australian show "Who Wants to Be a Millionaire." Which in turn is a copy of the CBS show "The $64,000 Question" from 1955, because you can't copyright show formats.
You also can't copyright procedures, rules or recipes. You also can't copyright an artistic school or style.
So, you are conflating the term "duplicate" with the term "copy". Duplicates are covered by copyright, as they are an exact reproduction of an artistic work.
A rabbi crawling out of a drain grate in the style of Studio Ghibli absolutely is not.
To take a photographic image and then redraw that into a new picture in an entirely different style is not a copy. It certainly isn't a duplicate.
To combine two images into a fusion of artistic representations into an image that has never existed before is not a copy.
You are correct. A convolutional neural network will not be able to generate images that match criteria outside its training data set.
But you can't speak Chinese without every having learned it. So what? What does that have to do with anything at all?
Your major complaint seems to be that Convolutional Neural Networks are not, in fact, people. You got me! They have less neurons than a flatworm. They are certainly less complex than the nervous system of a garden slug. They get things wrong and draw stupid shit all the time.
That doesn't mean they are a Xerox machine. They have a different function and are really good at certain tasks, especially of classification. They are very scalable. They are not thinking machines in any sense of the word "thinking".
Really?
Transformative is the standard. And it can be adjudicated. Pure copies by hand are called forgeries and are not legal to sell. I don't know why you think a hand copy is immediately substantially transformative when there's centuries of case law saying precisely the opposite and offering real tests to be used in making the determination.
If I can recognize the style then it's obvious you copied it.
No, you are, in an effort to argue in bad faith. The tool makes copies. Sometimes those copies are so bad they actually do violate copyright. That they can be transformative does not relieve them of being copies. This is obvious.
Your previous example seemed to rely on the fact you could combine a known character with octopus parts and call that a "new work." The point being, even copyright law, with all it's exceptions, would call this a violation. You can't draw Iron Man with squid legs and call it a new character.
Effectively they are. So are humans. The difference is humans are additionally capable of creating entirely new things. These toys cannot. They will not. There is no upgrade to make them capable of it. There is no obvious technological path from the current implementation to a new one which is equally capable.
This thing can only copy. To get back on point your 30 year timeline is bogus. If you understand this technology so well then you should see this clearly.
Guy, when you enter a prompt into a Convolutional Neural Network, you can't point to the work that it is coping; because it isn't copying a single work, or even a couple. The NN is relying on the complex interactions of neurons to identify patterns within its output and to adjust those to make them better match a gestalt of that training.
You know this. You tried to correct my original response to point this out to me.
When you duplicate an image with a Xerox machine, or even a person visually copying an artwork by looking at it, you can point to the piece being copied. "There it is, that is the original."
Duplication is prohibited under copywrite law; but style isn't. Drawing Spongebob Squarepants in the style of Vincent Van Gough is a creative, transformational work even if you can recognize both the character and the style.
There may be issues about selling or distributing that work, because Spongebob is covered by trademark etc., but there is no question about the transformative and artistic nature of the piece. Juxtaposing the vapid childrens' character and one of the most famous Dutch painters is, in itself, an artistic statement.
From https://www.clrn.org/can-you-copyright-an-art-style/
I get that you fucking hate Neural Network generated art. I understand, but you are just making shit up to win arguments. You are demonstrably wrong, which is worse because you know you are wrong, and you don't give a fuck.
Every day we make fun of Neo-Marxist fuckheads for making up their own facts, and here you are doing it to win an internet slap fight.
Even if you manage to declare victory by utterly distorting terms (a leftist tactic) you will have convinced no one of your point of view.
There are really, really good reasons to distrust Convolutional Neural Networks and Large Language Models. I've given a few very good ones in other posts in this topic. "lol; they are just Xerox machines!' isn't one.
Oh, and BTW, I've never given a 30 year prediction. That was someone else.