So an amount of this deals with scale and the copyright status of the prior art.
The real problem is that our definition of "plagiarism" did not anticipate, and was never intended to describe, anything even remotely like ML.
Isn't this like saying "cars are just doing the same as horse drawn carriages but faster, they shouldn't be regulated differently"? That "faster" changes the game completely. Similarly, many of the laws applying to physical actions can't be immediately transposed to electronic actions.
The current rules were made under the assumption that we're dealing with humans and "natural" laws. Like the fact that humans eventually die while an AI can live forever or transfer the entirety of useful knowledge to the next generation. A human doesn't have time to steal in a lifetime as much as an AI can do in mere minutes. A human can be punished for breaking the law, especially for repeat offenses, whereas an AI can't be even if it breaks the law more times than all human artists put together. These are all complete game changers, speed is only one of the many compounding factors. Scale changes things, a punch in the face isn't just "a gentle tap, but faster". An AI can suck in all of human knowledge and work with all of it at the speed of light.
> How do you think humans learn art?
Humans as a race have created art from literally nothing. Humans as individuals have constantly advanced it with original material. AI seems to be by definition derivative. Humans are encouraged to create something as original as possible and attacked if they borrow too much, AI in "encouraged" to copy as much as possible and as this discussion stands to prove, also to "borrow".
This at the very least invalidates the too simplistic view that AI is the same as humans but faster.
No, they didn't? Humans don't have any magic creative juices that allow them to come up with ideas out of thin air, that's silly. Ideas are created from our brain using our past experiences and knowledge, why do you think some of our earliest art is of animals?
Also, I'm not sure comparing artists to horse-drawn carriages is the best analogy to support your argument, considering horse-drawn carriages are almost nonexistent nowadays. I'd argue a better analogy would be something like chess or any other game. When computers beat humans at chess, nobody changed the rules of chess to make computers perform worse. People just acknowledged that computers were better.
In theory, you could have 1000 people watching cameras with a list of people to watch for in 1980, but in practice it was too expensive and not a consideration. Now you can cheaply record tens of thousands of hours of video indefinitely and run analyses on everyone in them. This has huge implications on freedom and society and calls for new rules.
Just because technology is doing the "same thing" that was done before at scale does not make it the same thing.
AI is the same.
> Ideas are created from our brain using our past experiences and knowledge
Humans experienced random things in life and some created art from it. Sometimes after hallucinogenic drugs they created art even from things that never existed. Then other humans created something new on top of the existing art. AI "experienced" already created art and was tasked to recombine pieces to create more art. Simply put, nobody gave an AI sounds of nature and had it come up with a symphony after some thousands of human years' equivalent of processing.
> Also, I'm not sure comparing artists to horse-drawn carriages is the best analogy to support your argument, considering horse-drawn carriages are almost nonexistent nowadays.
Not sure why this harms the analogy, it's meant to point out that "speed" makes a difference and changes the rules. Artists may also be nonexistent tomorrow. If AI is doing everything legitimately and creating art the same way as humans but faster and better, why would human artists continue to exist? And then what would you feed as training material to your AI? The dictionary definition of "love"? Sounds of nature? Art created by the previous AIs?
> When computers beat humans at chess, nobody changed the rules of chess to make computers perform worse.
Right, we just don't allow computers to play the game in official positions. Different rules, as it were.
for me art is not the end product but the process. I want to understand the artists intentions and often a work that looks blant, because deep because of the artist's explanations.
Take Agnes Martin's paintings as an example.
So what of the Dadaist, the postmodern and the rest, where art is only art because we say it is art, and the creation of art need only involve the “artist” scribbling their signature on the object to bestow it with “artistic merit”. I cannot see a good way to reconcile a philosophy (law is different here, plenty of art can be illegal, eg: graffiti) involving AI and art like this… Duchamp’s readymades, are the products of a factory’s meant for consumer use, are Art… AI art is the product of an sophisticated set of multiple software tools combined and then used by a person to produce their own artistic vision, the output is Art.
If Duchamp signing a urinal makes the urinal art then me using a prompt to craft a specific artistic outcome arguably required more creative effort to understand my “tools” and use them to produce a result that should also be considered art if I want to call it that.
I accept that we may come around and regulate and legislate these tools, just like how it’s illegal to make a graffiti masterpiece on someone’s wall without permission, it may eventually be illegal to use models trained with “unlicensed” data and the matter settled… but can we quit with this denigration of the output as somehow not actually art?
Regarding Duchamp's Urinal, the intention and the critique it voices is critical for being considered art (and a lot of people still argue about it). https://www.bbc.com/culture/article/20170410-the-urinal-that...
There have been a lot of upside down urinals in the world before Duchamp did that, yet nobody considered it art. Here again the intention matters. Matrix multiplications on big data don't have any intention.
However the AI process does not create a copyright (there is no human involved) and cannot use the transformative exceptions for fair use due to, again, no human being involved.
The thing to remember is AI literally copies the originals. While it mashes them together in a way that makes the lines hard to see that is in fact what it is doing.
Are they the best way to reward creation?
Do they serve humanity well?
In their entirety? No of course not.
Certainly the duration is ridiculous beyond measure.
But the idea of having no protections for your own work seems barbaric. At least for work that requires "work" for lack of a better word.
I think the laze faire nature of online memes for instance makes sense in that space.
> Are they the best way to reward creation?
I think that providing a mechanism for people to share their work without losing control of it makes sense.
> Do they serve humanity well?
While I wouldn't attribute societies recent online boon to them. I do think that having protections of this kind are one of the reasons we have seen such a huge investment in the space.
Like the current internet or not the hundreds of billions or trillions pumped into it would be harder to stomach if the internet were the wild west some would like it to be.
> a mechanism for people to share their work without losing control of it
Is impossible.
I see this argument a lot, but it cannot be true. The image database behind Stable Diffusion is 5 billion images resized to 512x512 each and compressed down to a roughly 100,000 gigabytes. But, you can run it on a model file that is as small as 2 gigabytes. The originals are not in there. They have not been compressed to 4 bits per image.
I've studied and used AI art generators for a long time now and I can tell you from both a technical and experiential POV, the images they create are not mashups.
20kb per image is like a jpeg worth of data, so certainly not throwing away anything.
A tiny model file without logic as to how you got to it isn't interesting. You could just throw away 99.5% (or whatever the ratio is) of the data and call it good. Mishmash data before you do and you might not even have to do that.
This style of AI is all about getting right past the human ability to perceive the mishmash and stop there. Given the underlying algorithm is literally about mishmashing I don't know why you are trying to fight that point.
Mishmashing implies that small clumps of pixels are in the model file somewhere. That's literally not true and has no relationship to how the algorithm actually works.
There's a ton of misinformation about how this works being tossed around. It's causing a lot of anger over ideas that simply false. That's what I'm fighting.
Don't quote the 2 GB one that one is disingenuous.
That is like copying the dictionary heap sorting by word frequency leaving only the most popular words and saying "see I couldn't have copied the dictionary it is too small to have done that".
And no the fact that it has "all" the data isn't material here.
As another commenter already pointed out, this is literally not possible. The model size for Stable Diffusion isn't nearly big enough to contain all its training data. Understanding the concepts (at least on some abstract level) of the artwork it's trained on is the only possible way a 6 GB[0] model can be based on such a massive data-set.
In conclusion, the important thing to remember is that AI does not just copy and paste the originals. It's simply not possible.
[0]: https://huggingface.co/stabilityai/stable-diffusion-2-1/tree...
I'm getting really tired of repeating this really simple concept that AI evangelists can't seem to understand.
'Digital devices are not afforded the same "observational rights" as humans. You can't just buy an extra ticket for a video camera and take it to watch Avatar with you. And you're definitely not allowed to sell those "observations" no matter how compressed the format.'
How do you think humans learn art? Humans aren't magic, they learn to make art by looking at and being inspired by other art, and I'd bet quite a lot of that art is copyrighted. AIs are just doing the same thing but way faster, and I don't see any issue with that. (Same with GitHub Copilot for that matter.)