Copying a work itself can be copyright infringement if it’s very close to the original to the point people may think they’re the same work.
We might be able to argue that the computer program taking art as input and automatically generating art as output is the exact same as an artist some time after general intelligence is reached, until then, it's still a machine transformation and should be treated as such.
AI shouldn't be a legal avenue for copyright laundering.
1) the artist is not literally copying the copyrighted pixel data into their "system" for training
2) An individual artist is not a multi billion dollar company with a computer system that spits out art rapidly using copyrighted pixel data. A categorical difference.
> automatically generating art as output
The user is navigating the latent space to obtain said output, I don't know if that's transformative or not, but it is an important distinction
If the program were wholy automated as in it had a random number/words generator added to it and no navigation of the latent space by users happened, then yeah I would agree, but that's not the case at least so far as ml algos like midjourney or stable diffusion are concerned
On 1, human artists are copying copyrighted pixel data into their system for training. That system is the brain. It's organic RAM.
On 2, money shouldn't make a difference. Jim Carrey should still be allowed to paint even though he's rich.
If Jim uses Photoshop instead of brushes, he can spit out the style ideas he's copied and transformed in his brain more rapidly - but he should still be allowed to do it.
(That's as opposed to a large language model, which does memorize text.)
Also, you can train it to imitate an artist's style just by showing it textual descriptions of the style. It doesn't have to see any images.
Going painting > raw photo (derivative work), raw photo > jpg (derivative work), jpg > model (derivative work), model > image (derivative work). At best you can make a fair use argument at that last step, but that falls apart if the resulting images harm the market for the original work.
They probably aren't doing that. Studying the production methods and WIPs is more useful for a human. (ML models basically guess how to make images until they produce one that "looks like" something you show it.)
Automated transformation is not guaranteed to remove the original copyright, and for simple transformations it won't, but it's an open question (no legal precedent, different lawyers interpreting the law differently) whether what these models are doing is so transformative that their output (when used normally, not trying to reproduce a specific input image) passes the fair use criteria.
But currently, first, there is a reasonable argument that the model weights may be not copyrightable at all - it doesn't really fit the criteria of what copyright law protects, no creativity was used in making them, etc, in which case it can't be a derivative work and is effectively outside the scope of copyright law. Second, there is a reasonable argument that the model is a collection of facts about copyrighted works, equivalent to early (pre-computer) statistical ngram language models of copyrighted books used in e.g. lexicography - for which we have solid old legal precedent that creating such models are not derivative works (again, as a collection of facts isn't copyrightable) and thus can be done against the wishes of the authors.
Fair use criteria comes into play as conditions when it is permissible to violate the exclusive rights of the authors. However, if the model is not legally considered a derivative work according to copyright law criteria, then fair use conditions don't matter because in that case copyright law does not assert that making them is somehow restricted.
Note that in this case the resulting image might still be considered derivative work of an original image, even if the "tool-in-the-middle" is not derivative work.
Say it with me: Computer algorithms are NOT people. They should NOT have the same rights as people.
And the weights. The weights it has learned come originally from the images.
Also, a jpg seemingly fits your definition as “no creativity was used in making them, etc” but clearly they embody the original works creativity. Similarly, a model can’t be trained on random data it needs to extract information from it’s training data to be useful.
The specific choice of algorithm used to extract information doesn’t change if something is derivative.