Scale, really. If you mix two tracks, it’s going to be pretty obvious that you have mixed two tracks. If you mix 15,000 tracks, no one is listening and saying “Hey, that’s just Singer-Guy’s track from PopularSong!”
I think the record company lawyers would beg to differ here.
Just because it’s harder to identify the source that was copied doesn’t mean its not copying
Maybe. I doubt there’s any cases of the record companies coming after someone who mixed 15,000 tracks to really make that argument though. You’d have a really hard time proving that your work was used to any degree significant enough to count as infringement.
Even in popular cases where Artist A sued Artist B for lifting a track or composition, there’s a lot of argument over what counts as use or what’s actually unique and what’s just standard popular music melodies, etc. And that’s just for one artist. Multiply by thousands and I don’t really see it being provable that one person’s work was significantly used.
The various law suits working their way through the legal system about AI generated art are exactly that
I think you’ll need to show a legal precedent that there is some magic number of original works you need to copy such that, when exceeded, your work stops being a derivative work.
Not really. Those are about the training, not the output. The argument isn’t “User Bob unfairly made a work that used part of my work”, it’s “AI Company unfairly used my work to train their model”. The lawsuits are against the AI training operations, not against the end users creating “derivative” works.
No one is trying to prove that the song I made using AI is 0.15% Captain & Tennille so I’m violating the their rights. The argument is that the model used their music without compensation in its training even if that model is thrown into the ocean and no one ever uses the data to create music.
As I said, cases get tied up for years over a single track. I don’t see it being a case if you have thousands of tracks. Feel free to disagree but I’ll believe it when I actually see it happen.
Nope that’s exactly what’s being claimed (among other things). Eg:
Vicarious Copyright Infringement (17 U.S.C. § 106): because the output of the LLaMA language models is based on expressive information extracted from the plaintiffs’ works, every output of the LLaMA language models is an infringing derivative work.
…and was dismissed.
Meta has moved to dismiss all claims except the one alleging that the unauthorized copying of the plaintiffs’ books for purposes of training LLaMA constitutes copyright infringement. The motion is granted. The remaining theories of liability, at least as articulated in the complaint, are not viable.
[…]
With regards to the vicarious liability argument, the complaint offered no allegation of the contents of any output to support the statement that “every output from the LLaMA language models constitutes an act of vicarious copyright infringement.” To prevail on a theory that LLaMA’s outputs constitute derivative infringement, the plaintiffs would have had to allege and ultimately prove that the outputs incorporate in some form a portion of the plaintiffs’ books.
That dismissal goes back to what I was saying: There SO MUCH information in there that proving that YOU specifically are being infringed upon with the output is very difficult.
That case, for now, in that jurisdiction. Others have been more successful:
I’m not saying that this will win out (especially in the current political environment 
). But that fundamentally is question that’s being answered. To me it’s a no brainier, it clearly makes it a derivative work. But it wouldn’t be the first time the courts come to what is in my opinion a completely nonsensical conclusion on a matter of technology and law.
You asked how this differed from someone mixing two tracks and my answer is the same: Because when you’re mixing thousands of tracks, someone is not going to be able to actually point to how it’s derivative of THEIR work and thus has no case. Which was the crux of the dismissal. Unless you can point and say “This part was mine” you essentially have no case and that’s the primary difference versus mixing two tracks.
In simpler terms, you don’t have a case for copying if you can’t actually say A is a copy of B.
This is same as saying taking all the fractions of cent from all the millions of bank transactions that happen every year isn’t stealing, as it would be very difficult each individual customer to identify how much you stole from them. That’s a nonsensical argument the difficulty of proving your guilt makes no difference as to whether it’s a crime or not
So in your interpretation of the law when does it cease to be a derivative work? When you copy 10 different tracks? 20? 100? 1000? 1000000?
No it’s not because cent fractions are a material thing and you can prove their loss. I can literally say “In this transaction, 0.005¢ went from this account to this account” and prove it. You can’t take a work made up of thousands of other works and say “On this page, this bit came from this other bit” and have the same evidence-bound case.
Beats me. Don’t care really. More than two, less than a bajillion. It’s irrelevant anyway because the question isn’t how many works are there total, it’s whether or not you can prove that Work A unfairly used your Work B. The more works wrapped in A, the harder the case for B until it becomes impossible.
One difference that I think is important is that the endgame of a human listening to music to be inspired by their heroes is almost never to outcompete those inspiration sources to death.
A person studies art to become an artist - to join the community of those artists and to hope to contribute their own art to the genre.
AI does it to displace the artists and pollute the genre with poor quality, high-volume commercial garbage.
This is the “art vs content” debate.
With content, the producer is subordinate to the consumer. You first determine what the consumer wants and then work to fulfil that desire. If it succeeds, then a content is fit for purpose, and therefore, “good” and if it doesn’t, then it’s not fit for purpose and “bad”.
With art, the consumer is voluntarily subordinate to the producer. The producer first determines what they want to say and it is up to the consumer whether they are interested in engaging but from a baseline desire to want to receive that message. If the message is legible, then art isn’t necessary, the producer can simply just say it flat out. But so much of what we say can’t be said with plain words, hence, art. Art is “successful” if the transmission of the message from the producer to the consumer is coherent and the message is one that provokes some kind of introspection or new insight in the consumer. Art is “unsuccessful” if the message is muddled or the message is trite or uninteresting.
Engaging with content as art or art as content is the root source of much confusion.
I’ll use food as an example. Your average burger is content, the goal of it is you want something delicious to make you happy. You’re judging it from purely selfish criteria, does it taste good, is it fairly priced, is the service good, are the restrooms clean, is the location convenient, the seller only exists insofar as to fulfil your selfish desires. Fine dining is art, the nutritional value and taste are only there incidental to a statement the chef wants to make via the experience they give you. The food can give you new perspectives on sustainability, or labor rights, or the technical possibilities of food etc. You are deciding whether you’re interested in what the chef wants to convey, how well they executed it and whether you appreciated it but fundamentally, if you go into a fine dining restaurant expecting solely deliciousness on your terms, you’ve engaged at the wrong level.
AI can create content, but it can only ever be a tool in creating art because content is in conversation with content and art is in conversation with art.
Imagine an AI burger restaurant, you could have AIs develop the recipes and robot AIs in the back cooking and AI could take feedback into account to continually optimize the recipe and even personalize the burger for each individual based on data gathered from your behavioural history. AI isn’t there yet but it’s conceivable for there to eventually be an AI burger restaurant. More importantly, a burger restaurant is convergent, you slowly approach the platonic ideal of a burger that is a roughly fixed point over time and space.
Imagine an AI fine dining restaurant, you can’t because the concept is incoherent. Art comes from a point of view that is in conversation with another point of view and we engage with the art because we’re interested in engaging with the creator behind the art. More importantly, fine dining is divergent, what was fine dining becomes less and less fine dining over time as it ages out. For something to be fine dining, it needs to be new and fresh and add something new to the conversation. Fine dining exists only in the context of the time and place it is in because it’s in conversation with the other art that is happening at the same time. A fine dining restaurant can, for sure, use AI the same way many fine dining restaurants use other cutting edge molecular gastronomy technology as a part of their art, but it can never be a solely AI restaurant the same way a solely AI burger restaurant could be.
On the side against the AI boosters, I’ll say that most AI evangelists aggressively misunderstand art, somewhat out of ignorance and somewhat for malicious purposes and their deliberate smearing of art vs content is engineered confusion in order to overstate what AI can do.
On the side against AI naysayers, I’ll say that I think we probably significantly overestimate the true demand for art vs content and that we’ve probably been enjoying an artificial subsidy the same way newspaper feature writers overestimated their importance to the business value of news vs classified ads that was only exposed when the entire enterprise was unbundled.
IMHO, the essential reading on this debate that informed a lot of my above thinking is Ted Chiang’s essay Why AI isn’t going to make Art which many people seemed to have misunderstood so my big wall of text is intended as appropriate context to then go read Ted’s piece.
Or maybe they just don’t give a fuck what some gatekeeper tells them is “art” and what isn’t. (Typically the same person who will call a banana taped to a wall real art.)
I wonder what the conversation about AI would be if people never applied the word “art” to it in the first place, since that is where the hangup seems to be. What if it had always been called “illustration”? Would we have indignant people demanding that AI illustration isn’t illustration? Or how about simply calling it AI image generation. Would angry people insist that it isn’t an image?
How about we dispose of the loaded term “art” entirely, come up with a new word for it. Call grids of pixels generated by AI “gleebzooks”. That way people can be allowed to like their gleebzooks as much as they want without someone snobbishly telling them that it isn’t a real gleebzook.
Thanks for willing to be such a model example for me!
This is the central point, and thank you for distilling it so well.
To put my own spin on it: saying AI creates art is like saying a paintbrush creates art. Instead, they are both tools people use to apply color to a canvas.
I’m also reminded of the thesis “the medium is the message”. AI exemplifies this perfectly. Some people want to use AI as the medium to express their own message. But other people see the medium (AI) as the message. So we get the discussions about what the message really is.
You are very welcome and I wear it as a badge of honor.
I’ll even concede to you and Ted. AI generated pixel grids are not art. There, you win!
But I’ll continue to enjoy making gleebzooks and seeing other well made gleebzooks. That definitely aren’t “art”.
Good enough for you?
FWIW, Maurizio Cattelan’s Comedian is trivially art because it simply cannot be engaged with as content. The very act of engagement with the piece requires subordination and the mistake of engaging with it as content is a part of the art.
In order to engage with Comedian, you have to, at a bare minimum, know who it was by, know what other pieces he’s done, read his artistic statement, be aware of basic history of the piece such as the banana you’re referring to is the 3rd edition he sold and his original piece was in 2019, have some basic understanding of the other pieces he’s in conversation with and how he fits into art history.
Once you’ve done all that, you can begin to assess whether his piece is “successful” or not (content is good or bad, art is successful or unsuccessful) and, more importantly, join in the conversation with others who have engaged with the piece and have varying, subtle and nuanced views on the matter.
If you don’t want to do all of that, that’s fine, that’s why art requires the act of voluntary subordination. But the subordination is inseparable from the art and the typical childish, defiant attitude is simply a failure to engage.
Works like Comedian necessarily require a large amount of effort to engage with which is why it’s so rarefied and alienating to a lot of people but that’s why it’s part of the market that’s the most “fine” of “fine art”.
I often try to use terms like “image generation” or “renderings” instead of “art” but, hey, “art” is only three letters to type and people like to humanize things so it’s probably a losing battle regardless. Pretty much like people who cross their arms and say “that’s not REAL artificial intelligence” as though anyone else cares or is going to stop using the term. Battle’s lost, go home.
As I said before, I don’t care if it’s art or not. An image – made by who/what ever – doesn’t have to be “art” to make you think that’s pretty or neat or funny or cool or inspiring or makes you smile. It can have value to you regardless of where it came from.
I’m also not really sure if AI is much of a barrier to creating art. If I think that an image of a baby crawling down some train tracks is symbolic of the human condition, the art is in that concept. Whether I have an AI generate that image or whether I get a baby and put it on some tracks and use a camera or if I paint it, etc isn’t what makes it art. It might help make it more (or less) successful art due to its presentation and how people can access/experience it but the “art” is in the concept, not in how I arranged the studio lighting and rented a baby.
And AI has been used that way as a term of art (term of gleebzook?) in computer science since the 1950s, long before most of the people complaining that real AI is C-3PO were born.