(The difficult part is telling whether AI generated stuff infringes on the copyright of stuff in the training data. And the fact the onus is on the end-user of AI tools sucks...)
However - if you're arguing that copyright law should be extended to prevent use in training - that's a much tougher sell and one where I'm not 100% sure of my position.
What is perfectly clear are two things:
1. Github Copilot will sometimes produce non-trivial inputs it was trained on as outputs
2. Those outputs will not conform to the license
That is, in my mind, a clear violation. This could easily be solved by making training (or transformation or whatever) explicit in the license.
Edit: I single out Github copilot because such examples were found. I'm not sure you can find such examples for things such as Stable Diffusion or DALL-E, but the operating principle is the same.
DALL-E 1 will sometimes replicate watermarks (without being promted for it [1]). I'm not sure anyone has ever shown a 1-1 replication like what's shown with Copilot, but it's easy to get image models to create clearly derivative works (just ask for the Mona Lisa, or even a well-known modern work).
Training a model essentially creates a derivative work from the input data. GPL licensed code requires keeping the same license for derivative work. BSD-style licenses at the very least require retaining the copyright statement. It feels pretty clear to me that the trained model and code generated from it would have to adhere to the original license.
In a way ML is a bit like lossy compression. Would you think JPEG encoding and decoding an image strips the result of its original copyright?
Pretty clear indication that something is not yet 'transformative enough', IMO.
Same happens with Copilot and such - you can get it to spit out tens of lines of code verbatim pulled from someone else's project, with comments and all.
These kinds of things just don't sit right with me.
You're conflating two things here:
1. SD can generate art in the style of a specific person
2. It also adds signatures to some pieces
These are independent things. Sometimes it does 1 without 2 and sometimes 2 without 1. And the signature can appear on the wrong artist - or the signature is a weird mix of different signatures.
It's just that it's been trained on "paintings with signatures" so it includes them. It's not the damning evidence some claim it to be.
Comments and all would almost certainly be an infringement, as they are completely expressive and non-functional.
But you'll find that if you strip out those comments, it's possible that it would fail the test of the idea being separable from the expression [1].
[1] https://casetext.com/case/kern-river-gas-transmission-v-coas...
What is not fine is claiming that this is from the artist whose style is being copied. That makes your work a counterfeit.
To me, the legal protections for artists already exist: copyright offers protections against verbatim copying, laws against counterfeiting protect against original work made to be passed as if it were from the artist.
Whether these laws are still adequate in the face of AI-generated content is debatable and my guess is that we are in for interesting tests and debates around the subject.
Artists should have the means to protect their work. It's a condition to create a society where art can flourish in public. Everyone benefits.
However, artists should not be able to suppress the means of creation. Technology has always been a boon to Art. It can be the cause of huge shifts and some types of art may disappear in favour of new ones, but it's never hindered our innate need for creating art.
That creates a derivative work, which has a separate copyright on the creative addition (the moustache).
Training an AI has been found to be not creative work. The AI weights, as a result, are not a derivative work, and likely cannot be copyrighted.
Your GPL example is an unexpectedly poignant one, as well, though it will be to your chagrin.
If you publish your GPL code, and I learn from it and produce code that does the same thing, that is not restricted by copyright. Furthermore, if it's found that the expression of the code itself cannot be separated from the function, the expression is not copyrightable in the first place, and the GPL is redundant.
It's only in the instance that I reproduce your code exactly and the function is separable from the creative expression that your copyright is valid.
Regarding your JPEG compression, it's possible that a ML model will produce outputs that, of themselves, are copyright infringement. A recognizable picture of Ariel, for example. That doesn't mean that the training, the model, or anything else it produces is infringing.
The Mona Lisa has been in public domain for several centuries, so it's a pretty bad example.
https://en.wikipedia.org/wiki/Derivative_work#Examples_of_de...
Example: If I put my book on mybook.com, and you download it (legally), you can read it, learn from it, and produce works in a similar style, all without my consent, and copyright offers no tools to restrict that.
The only tool copyright offers to protect against that is distribution.
Since copyright doesn't govern your brain's "wetware", the comparison to human behavior is irrelevant.
The law has a very broad definition of creativity when a human is involved, even a human taking a picture of the Mona Lisa is considered creative.
Since humans train the A.I., and it takes creativity to design the training scheme, the A.I. would appear to me to clearly be a creative work and copyrightable.
You brought up a seperate issue of whether the model infringes the original work. And whether it's okay because the model doesn't resemble the original.
I don't think that's precisely the accurate legal standard. What's relavant is whether the copying into the model constitutes fair use.
One of the considerations of fair use is whether the use is transformative. But that's only one consideration a court will look at in determining whether the A.I. company can succeed in a fair use defense. I don't believe this is settled law.
But if there's any broad point I want to make, it's that the law does not ever consider a machine or piece of software "like a human" or say "If it's okay for a human to do it it's okay for a piece of software to do it."
Section 313.2 of the U.S Copyright compendium states, "The (copyright) office will not register works produced by a machine or mere mechanical process that operates randomly or automatically without any creative input or intervention from a human author".
Moreover, the Copyright Act's definition of a computer program is “a set of statements or instructions to be used directly or indirectly in a computer to bring about a certain result”. Neural net weights are clearly neither statements nor instructions, so it's unlikely the weights would gain copyright as a computer program.
It's indeterminate whether the SSO doctrine[1] is operable on neural net weights, as it can be difficult to determine in which order weights are used. If those weights were to gain copyright protection, they may not be enforceable on similar sets of weights.
Regarding whether resembling the original work is the legal standard for being considered an infringing work: the test applied by a court is called Substantial Similarity[2], and it's extremely unlikely that neural network weights would be substantially similar to training data when considering the idea-expression divide.
Copying into the model... are you suggesting that if I have an image that's represented by an array in memory, and I copy that array from one memory location to another in my pc, that act of copying is the infringing act?
Fair use may even be a bit non-sequitur in this case. If I look at copyrighted art, I don't have a fair use case, because I haven't incorporated the art into anything, though the act of viewing has changed my brain. Training a neural net on training data may not even be considered incorporation.
On your last broad point... okay? I suppose.
[1] https://en.wikipedia.org/wiki/Structure,_sequence_and_organi... [2] https://en.wikipedia.org/wiki/Substantial_similarity
You might want to read this article:
https://www.law.cornell.edu/wex/fixed_in_a_tangible_medium_o...
Your brain is immune from copyright lawsuits as being a derivative work because your brain is not a tangible medium of expression.
Your thoughts are not derivative works under the copyright act because your thoughts, seeing as they are not contained in a tangible medium of expression, are not works at all.
Computers are clearly tangible mediums of expression.
We can argue whether a particular piece of software is copyrightable or whether a dataset is protected under the copyright act's definition of compilations.
But it should not be up for debate whether whatever is happening in your brain when you learn something is in any way relevant in terms of copyright law.
I would also add the idea- expression dichotomy is irrelevent.
If the dataset was merely ideas like "A boy goes to Wizard school" then you might win on that ground. But if you feed it the complete works of Harry Potter, no court is going to claim that the data the machine created based on Harry Potter is an "idea".