| > Every human artist is trained on copyrighted works, reproduces that art during school, and often incorporates referential elements into their own art. The problem is that while some people view machine learning as equivalent to human learning (including some case law apparently) I do not believe most people see it that way. The way machine learning algorithms work today is decidedly very unique. A human artist learns to draw or paint a picture, but diffusion models are literally going from noise to final image directly. There is no sketching, or painting. If it generates artifacts that make an image look "drawn" it is because it "learned" what those artifacts look like. Even if a program really did paint something from scratch, and you sat there and directed it by typing in "over 600 prompts"... That's a lot closer to commissioning an artist than it is producing art. (Apparently, an artist who is very hard to communicate with, but thankfully has unlimited patience.) Photography is an interesting comparison point specifically because it is one of the lowest effort things that a human can possibly do to have copyright eligibility. I'm not saying photography is not a valid art form, but every picture you take is eligible for copyright equally even if it is not particularly creative. That's not the point of photographs being eligible for copyright, though. As flawed as copyright may be, one thing you can say about copyright is that it is very nuanced and specific, and the reasoning that led to the legal status of copyright in photography is not likely to be repeated for diffusion models. The legal status for photography seems to hinge on the fact that it clears the "fixed in a tangible medium of expression" requirement. This is a pretty complicated concept and I'm not a lawyer and won't try to conjecture on whether or not results from machine learning models used today could clear this bar. Either way, let's stop directly comparing "learning" for machine models that do things that humans literally can't and don't, such as "synthesize a fully-formed painting out of a bunch of latent noise" by analyzing trillions of pictures, to human learning wherein spend years and thousands of hours learning how to draw and paint (needless to say, analyzing factors fewer pictures in the process). They both behave very differently and accomplish very different goals. The only thing they really have in common is probably that they both boil down to statistics on some level. |
At 600 prompts, I think it's at least worth considering whether that similarity is just a superficial matter of the medium you're interacting with the tool in (text) as opposed to actual deficiency of creative input to the work.
Moreover, your actions can be exactly the same as directing a human yet still qualify you as the author of the work when that human is replaced by a tool - as with, say, a voice-activated camera.
> [...] the "fixed in a tangible medium of expression" requirement. This is a pretty complicated concept and I'm not a lawyer and won't try to conjecture on whether or not results from machine learning models used today could clear this bar.
Whether it's fixed in a tangible medium isn't the contentious term for generated outputs. That just means the the work is, for instance, written on a piece of paper - as opposed to being only a general idea in your head. It's well-established that the medium of digital images counts for this and, even if not, Allen's work had been printed out.