And yet, you have little to no understanding of art, but stand assured that this technology doesn’t infringe upon the artist’s practice or livelihood in any way. Perhaps you should get an MFA in order to understand this question all the way around.
1 Like
If this happens to be more than a fad this time, yes, it’ll affect livelihoods.
I read the F article and … I see a lot of basic misunderstandings of the technology in their public-facing site. I hope, for their sakes, that their legal paperwork is in a better format because otherwise they’ll lose on the basic facts of the case.
My understanding of art isn’t relevant to whether B was condescending in saying that artists don’t understand the technology. Which I’ll point out he said in the same sentence as “technologists don’t understand the law”.
Yeah, I have a family member who is a lawyer/professor specializing in Intellectual Property, and I’ve talked to him extensively about A.I… While the law isn’t totally settled, the current guidelines are VERY remix friendly. If you “significantly alter” existing art, there’s no copyright protection, and obviously anything made in Midjourney, etc. is wildly altered from the sources.
I really don’t see a way of putting this genie back in the bottle.
Which is kind of a nasty Catch-22. Those interpretations of fair use are there because the intentional recontextualization of media can be very effective in commenting upon and critiquing said media. But it’s that same fluidity that’s going to allow corporations selling auto-generated artwork to profit, probably. In the end it’s that situation where we’re once again devaluing the human in favor of some capitalist money mill. It’s a pattern that seems to be evergreen.
2 Likes
Well, an artist, even when aping a style, is creating an original work that is copyrightable in its own right, whereas the AI is algorithmically mixing up existing work, albeit in ways that are complex enough that humans can’t always tell which specific works are being used, not creating original, copyrightable work under current interpretations of the law. (It’s been previously established that for there to be a copyright for an image, the creator of the image must be a human being.)
2 Likes
I’m skeptical that the training process is as simple as just sampling word frequencies.
This article suggests that these programs will sometimes produce exact copies of the sampled works. This would not be possible if the original work were not encoded somewhere in the program.
An interesting experiment to help understand what these programs are doing might be to train the AI on only a single work of art. Would it just reproduce the art in whole or in part? If so then we could say that the AIs are really just doing a complex form of remixing.
3 Likes
If you trained something the size of stable diffusion/midjourney on 1 piece of art then it would very quickly reproduce only that. The only object in the universe as far as it is concerned is that one piece of art, and it will memorise it. It’s not a good experiment.
“Well enough” is a caveat in search of a context.
This technology is MSc-level comp sci. Most professional coders don’t understand it. Absent a settlement, this lawsuit could reel out over the decades to make SCO Group v. IBM look like an episode of Judge Judy.
6 Likes
Right here, we’re admitting that there is no originality possible in this current species of AI.
A human being, having seen no art ever in their life, handed a box of crayons, would still try something, experiment, play.
The equivalent human being would have to have been blinded, deafened, have no touch sense, no hunger sense…
…or be an infant.
Edit: Fine, an infant in the woods, if that helps.
Let’s take an easier example. An image classifier might ‘learn’ to distinguish images of cats from images of dogs. Reasonably simple example from 10 years ago.
All it knows is that any image it receives will be a cat or a dog. Give it a picture of a fish and it’ll tell you how confident it is that it’s a cat or a dog. The generators aren’t image classifiers, but it illustrates the point of how narrow an experience “give it one image” is.
An infant in the woods has already been exposed to far, far more stimuli than this. Even in the womb, they’ve received more stimuli.
There are a few big mistakes I see being made throughout the comment section.
The first is that copyright infringement requires direct copying. If I wrote and published an novelization of the new Avatar movie, I would absolutely be sued for copyright infringement, even though all the elements I used were, individually, uncopyrightable (e.g. character names, plot points). Every copyrightable thing is made up of smaller, uncopyrightable parts.
The second is jumping straight from the machine learning program to the user-requested artwork. Between this is the model developed by the machine learning program, which is itself a copyrightable work.
The question, at its heart, is whether this model is a derivative work of the training pictures (and the machine learning program).
To argue that it is not, the defendants would need to argue that the model is transformative fair use, not derivative. This might seem an easy claim. After all, an art generating program is very different from a picture, right?
All four factors of fair use are on the side of the artists here.
- The nature of the use is commercial, and it is not transformative in the sense that it “adds something new, with a further purpose or different character, altering the first with new expression, meaning or message.”
- The copyrighted works are creative, rather than factual.
- The entire image is used in the production of the new work.
- The effect on the market is obviously bad for the copyrighted work. We also need to point out that this factor includes hypothetical derivative works. There’s no reason an artist couldn’t license their work for an ML model.
3 Likes
Legally, it’s useful though. One argument it’s transformative is that the model only copies data about the work, not the work itself. If this data can reproduce the original work, then this argument becomes much weaker.
If we’re jumping straight to the final product then there’s a chance it falls on your side of the analysis, and maybe a chance it falls on the side of “a human didn’t make it, therefore there’s no copyright in it”. And if the derived work isn’t copyrightable then it can’t be commercial, maybe?
True, but it’s not been “trained on the work of professional artists.” It doesn’t have a degree.
The AI has exactly one learning capacity: visual, or rather information that its trained to stick together based on what we consider visually coherent. It understands, abstractly, a map of pixels and their assigned data.
The infant has every biological sense input, plus the fact of embodiment itself, the experience of living in the world. This is where the originality comes from. That emergent self. AI, in its current form, is nowhere near there, and I’m not convinced (yet) that it’s going to get there anytime very soon.
In the specific example of “tell a blank midjourney to look at a single work” you’d get over-fitting, which is something the full model has tried to avoid. It’s an example that tells you nothing about the full-size model, as long as the full-size model is relatively small compared to the training data.
Commercial just means its being created for profit. Whether it’s copyrightable is a different question entirely.
Whether something is copyrightable and whether it’s infringing are also separate questions. A work is only copyrightable once it’s fixed in a tangible medium. If I sell tickets to a live concert of other people’s work, it’s not copyrightable unless someone records it. It’s still copyright infringement, even if the performance, not being fixed in a tangible medium, is not copyrightable.
I was responding to the “gotcha” of “what if we trained it on a single image, and it reproduced that image exactly”. This diversion was on how a model trained on a single image is nothing like an untrained infant.
I agree with you that there is nothing similar between an infant with a full life experience and a machine learning model. That’s one of the reasons my first post in this thread pointed out how much I oppose calling it AI. Even these giant, giant models are less than the brain of a hamster, at best.
1 Like
Tangent, but: It still boggles me that these things can produce an image this coherent, and yet cannot get the number of fingers on a hand right. That seems like it should be the absolute easiest part of this image–it’s a finite number, there are billions of images of the human body to train from. Why does it make horror hands every single time?
