AI Art: Copyright and Creative Input
In the past few weeks, I have spent too much time reading arguments for and against "art"-producing algorithms. So much so that the thought of writing about it myself has crossed my mind once too often, so here we are.
I don't think many people hold the producers of copying machines responsible for copyright violations by their customers.
As someone relatively skeptical of the current implementation of copyright, it would be a step in the wrong direction, a hindrance of progress, to declare uncompensated inspiration a violation of copyright.
For humans, the default is opt-out. The owner of a work needs to protect it physically, with a patent, or with a contract to prevent others from using it as inspiration.
Should it be different for machines?
Personally, I find the practice of profiting from freely available work at the expense of the original creator distasteful, regardless of whether it's a function or a human involved. This is why I choose to apply appropriate licenses to my work.
I think it's useful to think of image generation models as lossy decompression algorithms. Given some compressed data X (and a seed), it produces an approximation of the original Y: `ลท = g(x [, seed])`. Comparing such a model to ZIP actually gives us a reasonable framework to think about things.
With any other compression method (e.g., ZIP), the algorithm used is irrelevant, and there is no distinction between X and Y when it comes to copyright. If Y is protected by copyright, so is X.
Setting the Stage
Before we dissect the core questions at the heart of this debate, let's establish a few foundational statements that I hope we can collectively agree upon.Should it be permissible to share functions capable of reproducing copyrighted works?
Yes. A tool capable of violating the law is not reason enough to forbid its use. For me, arguing that an "art"-generating algorithm should be restricted because it can produce images that are already copyrighted is unreasonable, considering that a simple copying machine is capable of that too.I don't think many people hold the producers of copying machines responsible for copyright violations by their customers.
Should a function be allowed to compete in historically human-dominated markets?
Yes. And it most likely will result in the loss of jobs. It is one of the features of capitalism that over time resources are allocated optimally (under some definition of optimal). If nobody is willing to pay for good G at price P, then producers unable to produce at these conditions have always gone out of business. It's nothing new: farming, manufacturing, even in art, countless jobs have either become entirely obsolete (copy-human) or the required workforce diminished (framers: 99% -> 2%) due to automation.The Crucial Questions
Now, let's zero in on the crux of the matter, which revolves around three pivotal questions:Should inspiration be compensated?
As it is right now, compensation for inspiration is not protected by law unless the influence of the inspiration is such that the produced work is not sufficiently different. The process used to produce the work is irrelevant in determining the extent of the infringement.As someone relatively skeptical of the current implementation of copyright, it would be a step in the wrong direction, a hindrance of progress, to declare uncompensated inspiration a violation of copyright.
Does the default allow a machine to use your work as inspiration?
The question is whether or not it should be opt-in or opt-out when it comes to permission to use a work for inspiration by a machine.For humans, the default is opt-out. The owner of a work needs to protect it physically, with a patent, or with a contract to prevent others from using it as inspiration.
Should it be different for machines?
Personally, I find the practice of profiting from freely available work at the expense of the original creator distasteful, regardless of whether it's a function or a human involved. This is why I choose to apply appropriate licenses to my work.
Is a prompt sufficient creative input?
Why not? A prompt is simply a compressed representation of the desired output.I think it's useful to think of image generation models as lossy decompression algorithms. Given some compressed data X (and a seed), it produces an approximation of the original Y: `ลท = g(x [, seed])`. Comparing such a model to ZIP actually gives us a reasonable framework to think about things.
With any other compression method (e.g., ZIP), the algorithm used is irrelevant, and there is no distinction between X and Y when it comes to copyright. If Y is protected by copyright, so is X.
Comments
Post a Comment