beetlejuice's ghoulish wifeI'm not that concerned about like rogue AI destroying humanity but I'm scared shitless of the human costs of a bunch of business geeks trying to shoehorn AI into every space of human endeavor as a hedge against workers (who they despise)
Something that really strikes me about people posting like "Hollywood is OVER!!!" over some weird uncanny valley nightmare video is - I genuinely do not believe these people understand that when directors or writers or illustrators have creative visions they want control over how those ideas are actualized. That the process of changing an idea into a piece of art is where the art COMES FROM
Clinical Rawlsian@mark The story emerging around "AI" has so many parallels to the history of the Luddites.
As @pluralistic says, "It’s not what technology does that matters, but who it does it for and who it does it to."
@mark @pluralistic The issue with AI isn't that it will automate some processes previously done by human creators. It's that it will do so using the output of those creators as "training data" without those creators ever being compensated.
And it will do so to further enrich people that are already rich at the expense of creators, performers, and audiences.
Steve Crist@MadMadMadMadRN @mark @pluralistic
The training process only ever uses existing work as a point of reference to determine if the computer is making correct predictions. Not even a single pixel of the original work is ever stored in the resulting model. It's very unclear why anyone would be entitled to compensation for that. It's like saying someone needs to be compensated for measurement. The goal is not to copy the original work, and that's the only thing IP protects.
Other people's work is an essential input to "AI". Without the data, the "AI" is useless. People who generated that data should be compensated for that.
There are already apps that render photos in the style of particular artists without paying those artists at all. This is not an academic problem we might encounter down the line. This is happening right now.
@MadMadMadMadRN @mark @pluralistic
This is a fundamental misunderstanding of the training process and the models themselves and how they work. As long as people have a shared concept of what makes a thing look "right," you could train an AI model on literally random noise as long as human beings provided it feedback like a rorschach test to say "sure, that sorta looks like a couch."
Existing work is simply a shortcut for this process. They are extant examples of what things look like.
But companies like OpenAI don't buy a bunch of couches and hire a bunch of people to take pictures and label those couches and upload those images to its servers. They just scrape the web for other people's work without compensating or even notifying them. They then fed those images into proprietary, closed systems that they sell to customers. At best, that's fencing the commons. To me, it's more like theft.
Calling it "measurement" doesn't change anything.
@MadMadMadMadRN @mark @pluralistic
Again, though, copyright does not protect an author from USE, it protects an author's right to COPY the image. I think it is an uphill argument that calculating probabilistic weights about what is or is not noise in an image, from millions of images of various concepts, is somehow infringing on an author's right to consent for replication. I do not see how that is anything other than a derivative work.
You may object to OpenAI, but I don't see the harm here.
@srcrist @mark @pluralistic I'm not going to pretend to be a copyright expert, but I'm pretty sure it protects use as well. That's why artists have to pay to cover a song written by someone else. It's why people can't make money by selling fan fiction of works not in the public domain.
But beyond the issue of copyright, the fact remains that private, for-profit companies are taking work produced by others without their consent or even knowledge and using it to make proprietary products.
@srcrist @mark @pluralistic Like, if I fed all of the code for Windows into an AI as training data and told it to make an operating system, do you think Microsoft would be cool with that? Do you think they would say I am just "measuring" their code to see what operating system code looks like?
@MadMadMadMadRN @mark @pluralistic
To be clear: if the algorithm did not duplicate any of Window's code, you have every right in the world to do exactly this. And this is even settled law. Emulators reverse engineer computer code every day to replicate it, and they are completely legal as long as they do not copy the code itself.
@MadMadMadMadRN @mark @pluralistic
You can replicate the function of code without replicating the actual text and syntax of the code. And this is legal. You cannot copyright an algorithm, and that's all a function in code is. So two people who understand the algorithm conceptually are free to write two different functions, in code, that lead to the same outcomes.
Andreas K@srcrist @MadMadMadMadRN @mark @pluralistic
The problem is that settled law for copyright is quite different depending upon the domain.
E.g. covering music can be trigger copyright, in some cases actually a couple of notes in a particular order, and the toll booth of the music industry is there to collect.
Fictional characters can be copyright protected.
https://en.wikipedia.org/wiki/Copyright_protection_for_fictional_characters
And so on. And because of WIPO, the perversions of 100+ countries copyright laws are suddenly relevant.
@yacc143 @MadMadMadMadRN @mark @pluralistic
It is for this exact reason that I would argue that the appropriate point of regulation with respect to computer generated images is at the point of output, and not training.
No amount of training makes Stable Diffusion or Midjourney capable of doing anything that Photoshop cannot already be made to do by a user. And the real issue here is in the output, as it is with any other image.
@srcrist @MadMadMadMadRN @mark @pluralistic I do admit that's my first instinctive PoV too, it's nothing more than a Photoshop/Gimp plugin.
@MadMadMadMadRN @mark @pluralistic
Your example is a replication. A cover is a rendition of the same song. You certainly do not have to pay an author to listen to a bunch of songs and compile a list of the songs with a particular chord progression, though.
All of this is an oversimplification in general, and IP law can be nuanced. But I don't think it's a very good argument to make that the sort of things "AI" training does are copyright infringement. If that's the case, so is a file hash.
@MadMadMadMadRN @mark @pluralistic
And, again, "take" implies that someone has been deprived of something. That framing is, in and of itself, wrong. They did not "take" anything. They ran an analysis algorithm on something and then the original was discarded.
@MadMadMadMadRN Because it is a copy? Pirated software is not some disambiguation of a copyrighted work. It's a copy of the copyrighted work. Your confusion of these two concepts is at the core of your misunderstanding.
@MadMadMadMadRN @mark @pluralistic
Your concerns about using AI to replicate copyrighted works are fine, but that's an output issue that has nothing to do with training the model. A model can replicate a style without ever having seen the work. It isn't able to replicate their style because it was trained on their work, it's able to replicate their style because that's just how diffusion works.
@srcrist @mark @pluralistic Honestly, I really don't care that much about copyright per se. I think a lot of these terms like "copy" or "use" or "rendition" are not immutable, universal constants but social constructs created by humans with tons of ambiguity and contradictions.
My concern is that OpenAI are using technology that would not be possible without the work of countless creators as inputs. That technology will be deployed in a way that makes the life of those creators worse.
@MadMadMadMadRN @mark @pluralistic
I mean, I guess it's fine, ultimately, to just be like "I don't care about these things, I think it should be illegal if it isn't already." But that's just an ethical opinion. The law very much cares about those terms.
And, again, those countless creators are only relevant in the fact that a concept must exist, yes, for a neural network to be trained. But the value is only in the aggregate weights for many, many examples of concepts. Not any specific work.
@srcrist @mark @pluralistic I never claimed to be issuing a legal opinion. You're the one who brought up copyright.
I guess I just don't see how "analysis in the service of making a proprietary product that people will have to pay to use" is fundamentally different from "use".
I don't want to criminalize the analysis of creative work in all circumstances. Like fair use protects certain types of copying and use, I think there could be "fair analysis" that protects certain types of analysis.
@MadMadMadMadRN @mark @pluralistic
I'm not criticizing you, to be clear, I'm just pointing out the distinction. All of this is effectively inextricable from issues of IP law anyway.
I would argue that existing concepts of fair use already permit the use of copyrighted images for training AI, generally, and there are a lot of legally tested parallel examples that support that.
There may be issues with OpenAI or StabilityAI's process specifically, but not at a 30,000 ft view, imo.
@srcrist @mark @pluralistic You clearly know more about copyright law than I do, so you're probably right. Though I have to admit I don't see how a computer program could analyze text, images, etc. without making a copy of them. But ,again, you know more than I do about how AI works.
@srcrist @mark @pluralistic For me, it's the is-ought distinction. Is it legal for ChatGPT to use the output of creators to make a system off of which OpenAI makes a lot of money without compensating those creators in any way? Maybe. Probably.
Ought it to be legal? To me, no.
@MadMadMadMadRN @mark @pluralistic
It is true that your computer even makes a copy of an image when you load a website. It does require a copy of the image in RAM in order to display it. But it's pretty well established that if you publish an image to be viewed publicly, the public's need to replicate the data in order to display it is not an infringement on your rights as an author. This is similar, imo.
It should also be noted that training uses destroyed copies of images, not originals.