I agree the definition is overly simplified. The main point was that training implicitly contaminates the model with a presumption of derivation, training is not some magic pixie dust you can sprinkle onto protected works and strip copyright away.
The next thing to discuss is if the derivative work is sufficiently transformative to be considered fair use without legal authorization from the rights owner. There is no easy analogy that can be made here, the type of derivation - transformation we talk about has no precedent in intellectual property law. My position to your challenge is that an AI text classifier/filter/recommendation engine is a sufficiently transformative derivation of the copyrighted works, whereas a general purpose machine that can produce works similar in style, content and character with the originals is not sufficiently transformative and should require authorization.
The way I propose we arrive at that conclusion, without any legal precedent, is one based on first principles, on the intent and final purpose of copyright law. It's a political philosophy position, namely that intellectual property is a social convention designed around the creation of a common good, it exists to promote "the progress of science and useful arts", general flow of ideas and knowledge, by creating an economic incentive for creators to produce and make public their work - as opposed to keeping it secret to control its distribution, or abandoning creation for other fields, both net negative outcomes that diminish the public good.
So when judging if the work of thinking machines is sufficiently transformative, we should ask: is what they output a net positive contribution to the common good of creation and widely available intellectual works, ideas or art? Is it at least not-negative? It's easy to make that argument with text classifiers, but much harder for something like Stable Diffusion. The algorithm it runs is completely dependent on human produced artworks, it cannot function without such an input and can't even produce a single creative bush stroke without them. Yet, the works it remixes using creative features of the originals can and indeed have already started to economically replace the work of original artists in the market place. So treating that derivation as fair use pushes society into a bad equilibrium, where artworks are less valuable and less likely to be produced, while the AI machine owner appropriates much of the economic value of the works it slurped in training. That's not 'fair use', and the AI machine as a whole is not 'fair use', even if some, or even all, of the works it produces could be considered taken individually, as fair use of the originals.
This will continue to be true for as long as human creators remain a key ingredient of the automated creation process. When and if a machine can start to paint after reading an university course on painting, then that would be indeed a fair use of those manuals.
The next thing to discuss is if the derivative work is sufficiently transformative to be considered fair use without legal authorization from the rights owner. There is no easy analogy that can be made here, the type of derivation - transformation we talk about has no precedent in intellectual property law. My position to your challenge is that an AI text classifier/filter/recommendation engine is a sufficiently transformative derivation of the copyrighted works, whereas a general purpose machine that can produce works similar in style, content and character with the originals is not sufficiently transformative and should require authorization.
The way I propose we arrive at that conclusion, without any legal precedent, is one based on first principles, on the intent and final purpose of copyright law. It's a political philosophy position, namely that intellectual property is a social convention designed around the creation of a common good, it exists to promote "the progress of science and useful arts", general flow of ideas and knowledge, by creating an economic incentive for creators to produce and make public their work - as opposed to keeping it secret to control its distribution, or abandoning creation for other fields, both net negative outcomes that diminish the public good.
So when judging if the work of thinking machines is sufficiently transformative, we should ask: is what they output a net positive contribution to the common good of creation and widely available intellectual works, ideas or art? Is it at least not-negative? It's easy to make that argument with text classifiers, but much harder for something like Stable Diffusion. The algorithm it runs is completely dependent on human produced artworks, it cannot function without such an input and can't even produce a single creative bush stroke without them. Yet, the works it remixes using creative features of the originals can and indeed have already started to economically replace the work of original artists in the market place. So treating that derivation as fair use pushes society into a bad equilibrium, where artworks are less valuable and less likely to be produced, while the AI machine owner appropriates much of the economic value of the works it slurped in training. That's not 'fair use', and the AI machine as a whole is not 'fair use', even if some, or even all, of the works it produces could be considered taken individually, as fair use of the originals.
This will continue to be true for as long as human creators remain a key ingredient of the automated creation process. When and if a machine can start to paint after reading an university course on painting, then that would be indeed a fair use of those manuals.