Multimodal Art
multimodal.art
multimodal.art
I confess my understanding of these tools is, at best, rudimentary, but they remind me a bit of sampling in music. In both cases, the source material may be radically distorted to arrive at the final product, but that hasn't mattered in the music biz—if you use it, you need to get it cleared, or you cannot monetize it.
Take the image of the dog with the red ball on this website. What if, in the data the AI was trained on, there was a usage-restricted image of a caramel dog in a grass field with a green ball in his mouth? Would the AI just use that image and change the color of the ball? Would it ignore that easy match and instead generate its own using however many other related images, resulting in something quite visually different from the restricted source image? Does it matter? If the AI-generated image is virtually identical to the source photo, or contains some piece of it fully copied and integrated into the new image (sampled), is that new image legally owned by the person who used the AI?
Can I train an AI on a data set consisting of a single image with a description of it, request an image of that exact description from the AI, and then publish and license the output as my own? If not, how big does a data set have to be before I can claim the output is novel and proprietary? Do a few blurry pixels or lossy compression artifacts prove an image has been sufficiently altered for new commercial use?
There’s precedent in musical sampling, visual collage and assemblage, and poetic cutups, but they don’t provide authoritative answers.
By the time the courts or regulators are ready move on the issue, the practice will be so established by big commercial efforts like Copilot that my money is on a fairly permissive approach if only because of inertia.
Yeah, I think you're probably right about this, for better or worse.
This is art. Generative algorithms are simply new tools to work in digital visual media. Fancy stencils. They're tapping into the same algorithms humans might use to generate things by hand, since they're function approximators trained in human output, but they're not ethically or legally more interesting than photoshop, to me, until there's a reason to question whether the tool or software is conscious. Until then, it's cleverly arranged math modules used to good effect.
> If not, how big does a data set have to be before I can claim the output is novel and proprietary?
There isn’t some exact N. That’s like asking how tall is tall. It’s fuzzy.
For sure. That's why I'm kinda surprised corporations aren't already trying to get a head of things and say any use is misuse, especially in the era of trigger-happy DMCA issuers.
Beneath a tinfoil hat, I'm inclined to think the IP holders most likely to issue these takedowns see the potential in low-to-no-cost content generation and want to be able to use this technology themselves. Much has been said about the pop music formula—how nice would it be to license software to generate hit singles instead of dealing with pesky creative types?
1. Did you see/hear/read the original?[0]
2. If so, does your work look like the original if you squint a little?
My gut feeling is that use of an AI with a copyrighted work in it's training set would constitute access for the purpose of determining copyright infringement. Substantial similarity would be determined on a case-by-case basis. It's trivially easy to get AI to spit out fragments of its training set; GitHub Copilot would infamously fill out the Quake inverse square root hack if you asked it to complete the comment "//evil floating point bit level hacking"[1].
There's no case law supporting any of what I just said; but I can't think of a counterargument to "training set = access". This creates a problem for anyone who wants to use AI to generate art and also follow copyright law. I have no fucking clue how you would go about clearing AI-generated art, other than to just pay whoever complains. This is actually how the creative industry works internally. Everyone just buys whatever they can license, tries not to do obvious infringements, and pays to make the problem go away. It's only the FOSS types like us that actually worry about being non-infringing, since we want to avoid copyright virality[2].
Also, I have a bit of a disdain for the current state of generative art. The people hyping up the technology want to use it to kill and eat artists, by replacing them with AI that can bury them in post-scarcity. In practice, current AI requires a lot of human effort and cajoling to corral it into creating something good. For every amazing demo of it spitting out an amazing work that matches the prompt, looks good, and is consistent; there's about ten or so failures where the AI just puts stuff wherever, never converges into an actual picture, or otherwise hands you back some kind of disgusting nightmare.
AI has no taste.
[0] There used to be 9th Circuit precedent for the "inverse ratio test", which says that if you have a lot of access, the judge or jury can squint harder when determining substantial similarity.
This has been overturned for obvious reasons, but not early enough to save Katy Perry from getting sued over an ostinato.
[1] Microsoft responded by... turning off Copilot if you entered in said comment. Not a good look.
[2] No, the GPL did not invent this problem, it's just a different strain of the virus.
> Also, I have a bit of a disdain for the current state of generative art. The people hyping up the technology want to use it to kill and eat artists, by replacing them with AI that can bury them in post-scarcity.
Yuuuuuup. Which makes me sad to say, because I had a big generative art phase when I was younger and loved seeing what I could do with hex editors and circuit-bending. I don't want to tread into the minefield of evaluating others' artistic processes, but my taste is for art with apparent intent and intentionality. In an era where notions of artistry have been largely divorced from technique, and technique is increasingly democratized anyway, I want to believe there is still some room for distinction between a neat image and a piece of art.
Many people see that as elitist gatekeeping or something, but if everything is valuable, nothing is valuable, and I think value is important.
I’m guessing the future isn’t here yet. It’s all hype. When someone says, my computer, ok. But this AI did it, not my fault. Like corporations are people—haha!
It kind of reminds me of http://www.infinityplus.co.uk/stories/blit.htm