3 ms·

“Whoever's using Photoshop to paint, however, does need to have seen existing artwork.” If this were true there would be no art, my friend.

hn_acker 3y ago
My bad. I failed to catch the assumptions that were in my head. Theoretically a painter can make art after having seen nothing but the real world or unedited photographs. Before people had the means to make and share paintings, they couldn't learn from paintings. They learned from the real world and their art styles. (One way of defining a painter's art style is "the opposite of what the painter can't paint".) Painters today generally don't learn how to paint by restricting themselves to looking at just the real world and realistic depictions, and the exceptions are not relevant to the question of what is copyright infringement. This is an assumption I didn't realize I was making, and now I am making it explicitly: The painter using Photoshop has seen paintings before (complete or in progress, doesn't matter), and the painter's creations are influenced by those past paintings. So I retract the following sentences from my previous comment: >> Whoever's using Photoshop to paint, however, does need to have seen existing artwork. >> A person using an ML model doesn't need to have seen existing works to ask for a generic output, but does need to have seen existing works to ask for an output similar to an existing work (outside of innocent name coincidences in the prompt like "a painting of a starry night"). And replace them with the following: In order to use Photoshop to produce a painting similar to an existing one, a painter usually needs to have seen existing paintings. There may be coincidences, such as when the painter's art style produces results similar to another painter's. In order to make a model produce an output similar to an existing work, the user usually needs to have seen existing works and subsequently give the model prompts referencing an existing work, author, or a characteristic/style strongly associated with a work/author. By strongly associated, I mean that a user who hasn't seen the Salvador Dalí painting "The Persistence of Memory" but has seen unrelated oil paintings might coincidentally give the prompt "oil painting of melting clocks" and get an output similar to that Dalí painting.
- voltaireodactyl 3y ago
  I really respect and appreciate your willingness to interrogate your own assumptions in public — it’s a beautiful way to be, and I hope I present similarly. My response would be that much hinges on the word “usually”, as in, the artist must “usually” have seen existing Dali work to recreate a similar aesthetic. While that’s certainly true, it’s also true that similar art evolves in different places at similar times. Which is to say, while “usually” is accurate, “always” is not. AI requires “always”. In the same way a photocopier, or a collage of existing works, requires the prior art to exist inside it in order to produce similar output. I don’t think it’s unreasonable to ask for compensation, if a tool is unable to produce the desired result without including your existing oeuvre of completed, copyrighted work.
  - hn_acker 3y ago
    > I don’t think it’s unreasonable to ask for compensation, if a tool is unable to produce the desired result without including your existing oeuvre of completed, copyrighted work. I agree, but I think there should be bounds to how much compensation the authors of the existing works are entitled to, because there is a well-established context in which relying on someone else's work doesn't entitle that person to compensation. Suppose that I review a book I bought legally and post my review online for anyone to see. I couldn't have come up with the review without reading the book. I would've been unable to produce my book review without putting the author's expression in the existing, copyrighted book into my mind. I don't need to claim that human minds and machine learning models are any more similar than the fact that they can take in information. Without specific information from a particular copyrighted work, I couldn't have made the book review. But as long as I legally obtained the information in the book (by buying the book), I don't need to pay the author for the book review itself because a book review likely is fair use. Due to the first sale doctrine, I don't owe the author anything for my book review even if I bought the book secondhand at a higher price than the original seller charged i.e. 1. I didn't compensate the original author for the book and 2. the person I bought the book from profited from selling the book to me. I've read the argument that the very scale of what an AI process can do (e.g. be trained on 1000s of books, can then summarize just as many books) means that humans using AI models should compensate authors in cases where a human manually performing the same kind of task doesn't have to compensate authors. That's an important argument, but figuring out the bounds for compensation with respect to scale is not clean either. Suppose that I make a training set consisting entirely of OCRed [1] scans I make of all of the books I own, including a 100 or so books that my friends gave to me. Here are a few versions of how I use the training set, in increasing order of complexity: EX1. I make a non-ML word counter, and the training is counting the frequency of each word across the sum of the books. I can can make the program predict the next word in a given prompt: pick a random word from the sum of the training set and the prompt; a word gets additional weight in the random selection according to the frequency of the word within the training set and also gets additional weight according to the frequency of the word within the prompt. EX2. Like EX1, except the program also takes into account the frequency of pairs of adjacent words in the text. EX3. Like EX2, except the program rates pairs of possibly non-adjacent words in the text according to frequency and the number of words between every two given words in the text. And so on. At a much later stage such as EX58, I switch to making machine learning versions of the program, which started as a word counter and became more of a word predictor. I think EX1 shouldn't require compensation to the authors of the books, but I'm not sure about EX57. How complex can the capabilities and accuracy (two separate but important factors) of a program trained on copyrighted works be until I need to pay compensation? There is definitely a difference between a sand grain (EX1) and a sand pile (EX??), but a law regulating AI has to consider where to place the boundaries to balance freedom of expression (which includes writing/sharing code and machine learning weights) and obligation to compensate authors. (I also think this is a problem beyond the scope of copyright. Copyright in the US doesn't reward the "sweat of the brow" [2] and doesn't apply in cases without substantial similarity, so I think AI compensation should be treated as completely separate from, not even ancillary to, copyright.) I don't have answers, but at the very least I don't want a hard line of "no AI model of any complexity without compensating authors". Predictive LLM that can write books (dubious quality, but possibly coherent)? The expected uses of such a model scale very easily in complexity and size, so the authors should get compensation, unless whoever makes the LLM only uses it for private personal purposes. A predictive LLM that can't remember more than 100 words at a time? Probably can't write a coherent book, so no compensation even if the training set includes entire books. [1] https://en.wikipedia.org/wiki/Optical_character_recognition https://en.wikipedia.org/wiki/Optical_character_recognition [2] https://en.wikipedia.org/wiki/Sweat_of_the_brow#United_States https://en.wikipedia.org/wiki/Sweat_of_the_brow#United_State...

hn_acker 3y ago

My bad. I failed to catch the assumptions that were in my head. Theoretically a painter can make art after having seen nothing but the real world or unedited photographs. Before people had the means to make and share paintings, they couldn't learn from paintings. They learned from the real world and their art styles. (One way of defining a painter's art style is "the opposite of what the painter can't paint".) Painters today generally don't learn how to paint by restricting themselves to looking at just the real world and realistic depictions, and the exceptions are not relevant to the question of what is copyright infringement. This is an assumption I didn't realize I was making, and now I am making it explicitly: The painter using Photoshop has seen paintings before (complete or in progress, doesn't matter), and the painter's creations are influenced by those past paintings. So I retract the following sentences from my previous comment: >> Whoever's using Photoshop to paint, however, does need to have seen existing artwork. >> A person using an ML model doesn't need to have seen existing works to ask for a generic output, but does need to have seen existing works to ask for an output similar to an existing work (outside of innocent name coincidences in the prompt like "a painting of a starry night"). And replace them with the following: In order to use Photoshop to produce a painting similar to an existing one, a painter usually needs to have seen existing paintings. There may be coincidences, such as when the painter's art style produces results similar to another painter's. In order to make a model produce an output similar to an existing work, the user usually needs to have seen existing works and subsequently give the model prompts referencing an existing work, author, or a characteristic/style strongly associated with a work/author. By strongly associated, I mean that a user who hasn't seen the Salvador Dalí painting "The Persistence of Memory" but has seen unrelated oil paintings might coincidentally give the prompt "oil painting of melting clocks" and get an output similar to that Dalí painting.

voltaireodactyl 3y ago
I really respect and appreciate your willingness to interrogate your own assumptions in public — it’s a beautiful way to be, and I hope I present similarly. My response would be that much hinges on the word “usually”, as in, the artist must “usually” have seen existing Dali work to recreate a similar aesthetic. While that’s certainly true, it’s also true that similar art evolves in different places at similar times. Which is to say, while “usually” is accurate, “always” is not. AI requires “always”. In the same way a photocopier, or a collage of existing works, requires the prior art to exist inside it in order to produce similar output. I don’t think it’s unreasonable to ask for compensation, if a tool is unable to produce the desired result without including your existing oeuvre of completed, copyrighted work.
- hn_acker 3y ago
  > I don’t think it’s unreasonable to ask for compensation, if a tool is unable to produce the desired result without including your existing oeuvre of completed, copyrighted work. I agree, but I think there should be bounds to how much compensation the authors of the existing works are entitled to, because there is a well-established context in which relying on someone else's work doesn't entitle that person to compensation. Suppose that I review a book I bought legally and post my review online for anyone to see. I couldn't have come up with the review without reading the book. I would've been unable to produce my book review without putting the author's expression in the existing, copyrighted book into my mind. I don't need to claim that human minds and machine learning models are any more similar than the fact that they can take in information. Without specific information from a particular copyrighted work, I couldn't have made the book review. But as long as I legally obtained the information in the book (by buying the book), I don't need to pay the author for the book review itself because a book review likely is fair use. Due to the first sale doctrine, I don't owe the author anything for my book review even if I bought the book secondhand at a higher price than the original seller charged i.e. 1. I didn't compensate the original author for the book and 2. the person I bought the book from profited from selling the book to me. I've read the argument that the very scale of what an AI process can do (e.g. be trained on 1000s of books, can then summarize just as many books) means that humans using AI models should compensate authors in cases where a human manually performing the same kind of task doesn't have to compensate authors. That's an important argument, but figuring out the bounds for compensation with respect to scale is not clean either. Suppose that I make a training set consisting entirely of OCRed [1] scans I make of all of the books I own, including a 100 or so books that my friends gave to me. Here are a few versions of how I use the training set, in increasing order of complexity: EX1. I make a non-ML word counter, and the training is counting the frequency of each word across the sum of the books. I can can make the program predict the next word in a given prompt: pick a random word from the sum of the training set and the prompt; a word gets additional weight in the random selection according to the frequency of the word within the training set and also gets additional weight according to the frequency of the word within the prompt. EX2. Like EX1, except the program also takes into account the frequency of pairs of adjacent words in the text. EX3. Like EX2, except the program rates pairs of possibly non-adjacent words in the text according to frequency and the number of words between every two given words in the text. And so on. At a much later stage such as EX58, I switch to making machine learning versions of the program, which started as a word counter and became more of a word predictor. I think EX1 shouldn't require compensation to the authors of the books, but I'm not sure about EX57. How complex can the capabilities and accuracy (two separate but important factors) of a program trained on copyrighted works be until I need to pay compensation? There is definitely a difference between a sand grain (EX1) and a sand pile (EX??), but a law regulating AI has to consider where to place the boundaries to balance freedom of expression (which includes writing/sharing code and machine learning weights) and obligation to compensate authors. (I also think this is a problem beyond the scope of copyright. Copyright in the US doesn't reward the "sweat of the brow" [2] and doesn't apply in cases without substantial similarity, so I think AI compensation should be treated as completely separate from, not even ancillary to, copyright.) I don't have answers, but at the very least I don't want a hard line of "no AI model of any complexity without compensating authors". Predictive LLM that can write books (dubious quality, but possibly coherent)? The expected uses of such a model scale very easily in complexity and size, so the authors should get compensation, unless whoever makes the LLM only uses it for private personal purposes. A predictive LLM that can't remember more than 100 words at a time? Probably can't write a coherent book, so no compensation even if the training set includes entire books. [1] https://en.wikipedia.org/wiki/Optical_character_recognition https://en.wikipedia.org/wiki/Optical_character_recognition [2] https://en.wikipedia.org/wiki/Sweat_of_the_brow#United_States https://en.wikipedia.org/wiki/Sweat_of_the_brow#United_State...