Suno & Udio To RIAA: Your Music Is Copyrighted, You Can't Copyright Styles (torrentfreak.com) 85
AI music generators Suno and Udio responded to the lawsuits filed by the major recording labels, arguing that their platforms are tools for making new, original music that "didn't and often couldn't previously exist."
"Those genres and styles -- the recognizable sounds of opera, or jazz, or rap music -- are not something that anyone owns," the companies said. "Our intellectual property laws have always been carefully calibrated to avoid allowing anyone to monopolize a form of artistic expression, whether a sonnet or a pop song. IP rights can attach to a particular recorded rendition of a song in one of those genres or styles. But not to the genre or style itself." TorrentFreak reports: "[The labels] frame their concern as one about 'copies' of their recordings made in the process of developing the technology -- that is, copies never heard or seen by anyone, made solely to analyze the sonic and stylistic patterns of the universe of pre-existing musical expression. But what the major record labels really don't want is competition." The labels' position is that any competition must be legal, and the AI companies state quite clearly that the law permits the use of copyrighted works in these circumstances. Suno and Udio also make it clear that snippets of copyrighted music aren't stored as a library of pre-existing content in the neural networks of their AI models, "outputting a collage of 'samples' stitched together from existing recordings" when prompted by users.
"[The neural networks were] constructed by showing the program tens of millions of instances of different kinds of recordings," Suno explains. "From analyzing their constitutive elements, the model derived a staggeringly complex collection of statistical insights about the auditory characteristics of those recordings -- what types of sounds tend to appear in which kinds of music; what the shape of a pop song tends to look like; how the drum beat typically varies from country to rock to hip-hop; what the guitar tone tends to sound like in those different genres; and so on." These models are vast stores, not of copyrighted music, the defendants say, but information about what musical styles consist of, and it's from that information new music is made.
Most copyright lawsuits in the music industry are about reproduction and public distribution of identified copyright works, but that's certainly not the case here. "The Complaint explicitly disavows any contention that any output ever generated by Udio has infringed their rights. While it includes a variety of examples of outputs that allegedly resemble certain pre-existing songs, the Complaint goes out of its way to say that it is not alleging that those outputs constitute actionable copyright infringement." With Udio declaring that, as a matter of law, "that key point makes all the difference," Suno's conclusion is served raw. "That concession will ultimately prove fatal to Plaintiffs' claims. It is fair use under copyright law to make a copy of a protected work as part of a back-end technological process, invisible to the public, in the service of creating an ultimately non-infringing new product." Noting that Congress enacted the first copyright law in 1791, Suno says that in the 233 years since, not a single case has ever reached a contrary conclusion.
In addition to addressing allegations unique to their individual cases, the AI companies accuse the labels of various types of anti-competitive behavior. Imposing conditions to prevent streaming services obtaining licensed music from smaller labels at lower rates, seeking to impose a "no AI" policy on licensees, to claims that they "may have responded to outreach from potential commercial counterparties by engaging in one or more concerted refusals to deal." The defendants say this type of behavior is fueled by the labels' dominant control of copyrighted works and by extension, the overall market. Here, however, ownership of copyrighted music is trumped by the existence and knowledge of musical styles, over which nobody can claim ownership or seek to control. "No one owns musical styles. Developing a tool to empower many more people to create music, by scrupulously analyzing what the building blocks of different styles consist of, is a quintessential fair use under longstanding and unbroken copyright doctrine. "Plaintiffs' contrary vision is fundamentally inconsistent with the law and its underlying values." You can read Suno and Udio's answers to the RIAA's lawsuits here (PDF) and here (PDF).
"Those genres and styles -- the recognizable sounds of opera, or jazz, or rap music -- are not something that anyone owns," the companies said. "Our intellectual property laws have always been carefully calibrated to avoid allowing anyone to monopolize a form of artistic expression, whether a sonnet or a pop song. IP rights can attach to a particular recorded rendition of a song in one of those genres or styles. But not to the genre or style itself." TorrentFreak reports: "[The labels] frame their concern as one about 'copies' of their recordings made in the process of developing the technology -- that is, copies never heard or seen by anyone, made solely to analyze the sonic and stylistic patterns of the universe of pre-existing musical expression. But what the major record labels really don't want is competition." The labels' position is that any competition must be legal, and the AI companies state quite clearly that the law permits the use of copyrighted works in these circumstances. Suno and Udio also make it clear that snippets of copyrighted music aren't stored as a library of pre-existing content in the neural networks of their AI models, "outputting a collage of 'samples' stitched together from existing recordings" when prompted by users.
"[The neural networks were] constructed by showing the program tens of millions of instances of different kinds of recordings," Suno explains. "From analyzing their constitutive elements, the model derived a staggeringly complex collection of statistical insights about the auditory characteristics of those recordings -- what types of sounds tend to appear in which kinds of music; what the shape of a pop song tends to look like; how the drum beat typically varies from country to rock to hip-hop; what the guitar tone tends to sound like in those different genres; and so on." These models are vast stores, not of copyrighted music, the defendants say, but information about what musical styles consist of, and it's from that information new music is made.
Most copyright lawsuits in the music industry are about reproduction and public distribution of identified copyright works, but that's certainly not the case here. "The Complaint explicitly disavows any contention that any output ever generated by Udio has infringed their rights. While it includes a variety of examples of outputs that allegedly resemble certain pre-existing songs, the Complaint goes out of its way to say that it is not alleging that those outputs constitute actionable copyright infringement." With Udio declaring that, as a matter of law, "that key point makes all the difference," Suno's conclusion is served raw. "That concession will ultimately prove fatal to Plaintiffs' claims. It is fair use under copyright law to make a copy of a protected work as part of a back-end technological process, invisible to the public, in the service of creating an ultimately non-infringing new product." Noting that Congress enacted the first copyright law in 1791, Suno says that in the 233 years since, not a single case has ever reached a contrary conclusion.
In addition to addressing allegations unique to their individual cases, the AI companies accuse the labels of various types of anti-competitive behavior. Imposing conditions to prevent streaming services obtaining licensed music from smaller labels at lower rates, seeking to impose a "no AI" policy on licensees, to claims that they "may have responded to outreach from potential commercial counterparties by engaging in one or more concerted refusals to deal." The defendants say this type of behavior is fueled by the labels' dominant control of copyrighted works and by extension, the overall market. Here, however, ownership of copyrighted music is trumped by the existence and knowledge of musical styles, over which nobody can claim ownership or seek to control. "No one owns musical styles. Developing a tool to empower many more people to create music, by scrupulously analyzing what the building blocks of different styles consist of, is a quintessential fair use under longstanding and unbroken copyright doctrine. "Plaintiffs' contrary vision is fundamentally inconsistent with the law and its underlying values." You can read Suno and Udio's answers to the RIAA's lawsuits here (PDF) and here (PDF).
Woah (Score:1)
Re:Woah (Score:5, Interesting)
Probably more than you think. Did you read the filing? They're not just defending what they're doing as fair use, but also going after the record labels for decades of unfair monopolistic practices, including ("on information and belief") imposing terms on services that if they allow AI content they can't have access to their music libraries, and so are seeking to have them disqualified for copyright abuse.
One thing I did find odd with their filing was as follows:
* Plaintiffs played a dishonest game of "Million Monkeys With A Million Typewriters" to insinuate that if you enter a prompt suggestive of a given song, you get a track similar to that song back.
* Defendants cited the Plaintiffs words that they are only alleging copyvio on the training data, not any outputs, and so wrote off addressing output similarities entirely in their filing.
But surely (A) it's still relevant, in terms of market impacts, even if it's not the primary point of the case; and (B) it would be a good way to expose them as dishonest and attempting to mislead the court.
Example: this track [youtube.com] was cited as evidence that, if you give the prompt "A famous 70s pop song about queens who dance, by a Swedish band that rhymes with fabba, europop, disco, keyboard, from an album that rhymes with jarrival", you get something similar to Dancing Queen. Except the creator (Ed Rex) omitted how many times he ran that prompt, and still had to crop it down to just four seconds to get his "similarity". This is what you get if you don't play A Million Monkeys With A Million Typewriters [youtube.com] with that prompt.
The other thing that they missed by not addressing this point was the ability to use the plaintiffs' own words against them. The RIAA filed an amicus brief in the Stairway to Heaven case arguing that a few seconds of relatively similar riffs isn't at all copyright infringement, nor a threat to and indeed if one used such a standard, it would entirely destroy music creation as we know it. But that's the very standard they pursued in this case. Citing the RIAA against the RIAA, page after page of debunking their claims with their own amicus brief, would have been, frankly, hilarious.
On the upside, they openly admitted that there "probably" are cases of the plaintiffs' music in the scraped training dataset, and that the plaintiffs will have full opportunity to sift through it and all records have been preserved for them. This isn't the sort of thing they should have been cagey about admitting; it's good for them that they were upfront and nonhostile about it rather than trying hide behind a shield of "proprietary information".
Re:Woah (Score:4, Interesting)
Believe it or not, a million-in-one chance of users deliberately trying to abuse your system in violation of your TOS succeeding through random chance in producing a couple seconds that sound kind of like a copyrighted song is not a legally actionable case against the developer. You can't sue Adobe if one of its users uses Photoshop to draw Donald Duck, and Adobe has zero safeguards in place to prevent you from doing so.
Even against said individual, a lawsuit would be highly unlikely to succeed (the widely panned and dubious Blurred Lines decision notwithstanding). The very reason why the RIAA filed an amicus brief in the Stairway to Heaven case (which, like most similar cases, the defendant won) is because if such a standard of "a couple seconds that sound kinda like something else" were allowed to stand, they could get sued for most of their catalog.
Most record companies don't even own the melodic rights anyways, just the recording (there are two components to a musical copyright, melodic and master recording). US copyright law on the topic: (9 17 U.S.C. 114(b)):
This is how, for example, Taylor Swift can get away with this [shepwedd.com]. It was the recording industry itself who sought this provision. They didn't want some guy selling copies of records out of the back of a truck, but didn't want to invite a lawsuit every time a new artist who signed with them released a song that sounded kinda similar to one of the millions of songs that some other artist had made.
It should also be added that if you can just randomly produce something that nominally "violates copyright", two things further come up:
1) You've automatically raised serious concerns about invalidating the copyright because the standards for something being copyrightable involve the work requiring human creative endeavour, not just some random sequence of notes / text / colours / etc, and if it's so short and simple that it can just be created randomly, it's highly unlikely you'll be able to defend it as being sufficiently creatively unique.
2) Copyright law requires access to the copyrighted material to be a violation. If two people independently create the same work, with neither having access to the other's material, not only is there no copyright violation, but they can both end up with a copyright on the exact same work (note: access to it, it doesn't mean having to prove that the person directly viewed the material).
And lastly, even if you did have access, didn't invalidate the copyright, and the melodic rights holder was defending their work, again, not only is this about you (not whatever software you used), but you can just get a mechanical license [wikipedia.org] for a 9-cent royalty payment. Cannot be turned down; it's automatic.
This will be an epic fight (Score:4, Interesting)
Hardcore commercial pirates against hardcore content mafia. Lets hope the annihilate each other
Re:This will be an epic fight (Score:4, Interesting)
We can only hope, but since the big investors are the same, you'll see how these lawsuits will eliminate the smaller "AI" fish that dilute the returns on investment, and leave the top dogs alone.
Of course, since it is a bit of an open question what part of the vomit by degenerative AI is "original" and "music", it will be mildly interesting to see how the arguments develop in court, but I predict that we will only see a "settlement" statement with no details on what the settlement actually is.
No side needs clear and unambiguous precedents here and now, as the "business model" of "AI applications" is still "evolving" ;)
Re: (Score:1)
Re: (Score:2)
Yes, I can. What does it have to do with the topic of this thread, though?
Re: (Score:1)
Re: (Score:1)
I'm afraid something is lost here because of the medium.
But yeah, we'll have an $1 general quantum computer in 2038, it will run off fusion energy too cheap to meter and give us the ultimate answer. And it will carry its own AI model, which will program it to do just anything.
Re: (Score:2)
Maybe, but both the hardware and algorithms currently are lacking [wikipedia.org].
You don't need to go that far, though. If we were to transition from classical backpropagation to predictive coding networks, we could probably implement those in analog on neuromorphic hardware and let physics do the calculations for us (e.g. multiplication times a constant is just a transistor functioning as a variable resistor; summation is just the junctioning of currents; etc). A key requirement of neuromorphic hardware however is that
Re: (Score:2)
You keep posting these two lines as if they are related to what is under discussion.
https://slashdot.org/comments.... [slashdot.org]
Yet they never are.
Nowhere is anyone arguing that "AI" is an existential threat to society, will destroy billions of jerbs, think of those poor starving artists while everyone rocks to ChatGPT tunes!" or whatever other fantasy you come up with. It is an objective truth that the output of the degenerative "AI" is garbage.
You must troll harder, or ask ChatGPT for an advice, maybe it could do be
One little problem with their argument (Score:3)
Gen AI is as close as we have seen in 233 to "something new under the sun" for the courts to consider. In 233 years they've never had to deal with a "back end process that consumes protected material, analyzes it and generates new content that competes in the marketplace with the original protected works."
I suspect the federal courts are going to have a serious problem with that, and I think we are more likely to see a major new exception to fair use than the free for all they're hoping for.
Re:One little problem with their argument (Score:4, Interesting)
It's a losing fight. There's lots of public domain and independent music, and lots of musicians who would be happy to play it for cheap, if human performance is even required.
I doubt the RIAA's licensing bans use in a cost function for a gradient descent optimizer either, so as long as someone was smart enough to use a licensed library they should be fine. If that someone is super lucky the RIAA might insert that into their licensing terms, the courts will uphold it, and they'll have a monopoly until somebody else recruits enough struggling songwriters to generate another dataset.
Or maybe the judge will say "you want to stop people from listening to music so they can make more? Get out of my courtroom."
Re: (Score:3)
Or maybe the judge will say "you want to stop people from listening to music so they can make more? Get out of my courtroom."
Or maybe the judge will say "you want to stop people from listening to music so they can make more? OK, fine by me." And he's upheld on appeal.
And the entire music industry dies a well deserved death.
Re:One little problem with their argument (Score:4, Interesting)
Replace music with software. Consider that software say, under a FOSS license like the GPL.
Then by that logic, a generative AI trained on GPL source code, may produce source code that is free of the GPL.
Is that really what you want in the end? Because that's an awesome way to lock up open source - we're not using Linux, we're using AI generated Linux workalike!
Copyright is required for F/OSS to work - without it, the F/OSS license is not applicable. It only works because the user is presented with a choice - use the software under standard "all rights reserved" copyright, or if you follow the license, you get additional rights to the software.
So as much as you might hate on the RIAA, it's also keeping F/OSS software around.
Re:One little problem with their argument (Score:4, Insightful)
That's fair and fine enough. Now dump the binaries of proprietary software into the AI and have it spit out a Free workalike.
And what of it if someone does dump the Linux source into an AI and cranks out a proprietary work-alike? Why would I choose that over the already well tested and Free one?
Re: (Score:2)
I have an example from real life. I bought some hair conditioner. I liked it. Six months later when I went back to the drugstore to get another bottle, I couldn't find it. What I found was 16 feet wide by six feet high bunkers of hair conditioner. Alllll different formulations, for thick hair, for fine hair, restoration oils, protection from UV, etc. For starters, I couldn't find the exact brand and formulation I liked, so I was forced to
Re: (Score:2)
I'm pretty sure it's going to remain easy to find your favorite Linux distro no matter how many crappy proprietary mee toos are out there. Just go to the same website,
Re: (Score:1)
Having spent the last few days trying to get ChatGPT to help me with the wave function collapse algorithm to use with 3d tile meshes in Blender, I can safely say that there is not much of a threat there of AI generated code pushing out human developers.
The code ChatGPT spat out was seriously flawed, and though it initially looked like it should work, had a propagate function that only propagated possible states using condition rules for the first possible tile in a cell, rather than all the possible tiles i
Re: (Score:2)
Oh yes, that would be terrible! A non-GPL UNIX-alike would be a terrible threat to Linux! Fortunately such a thing has never existed!
Wait, you could use that hypothetical model to make anything, GPL or not. Yeah, sounds terrible.
Re: (Score:1)
Re: (Score:2)
The funny thing is how foreseeable this all should have been.
A detector is inherently a generator. If you can take a dataset and detect how well it matches X, then you can also figure out gradients to adjust the dataset to more closely match X. So you can start with random noise, and end up with something that's fully X, to within the quality limits of your detector.
Re: (Score:2)
Re:One little problem with their argument (Score:5, Insightful)
In 233 years they've never had to deal with a "back end process that consumes protected material, analyzes it and generates new content that competes in the marketplace with the original protected works."
Bullshit. That is what all artists have done for as long as protected material has existed. From the time the first cave man draw a mammoth on a cave wall with soot from his torch, that is the only way it's been done, for longer than recorded history.
The only difference is that it's been automated, and that's not something that copyright law prohibits.
Re: (Score:2)
Re: (Score:1)
Except people actually can come up with something outside the bounds.
I agree that people can, but I disagree that computers cannot.
Computers (LLMs) can only interpolate. That is the math of what they are able to do.
For current LLMs, maybe, but for computers in general, not necessarily.
Introduce randomness (ideally using a source of true randomness to avoid arguments about predictiveness) and you can easily go beyond interpolation.
Re: One little problem with their argument (Score:2)
Re: (Score:1)
Re: (Score:2)
Only some people can. Most people aren't even able to interpolate.
Re: (Score:2)
I'm not sure you understand how large the latent space is.
What's being thrown out by reduction from output space to latent space is incoherent things. E.g. one bit of random static is basically indistinguishable from a different one, so there's no need for separate positions in the latent space to recognize them. But the spaces are still vast. LLMs for example cover "basically every concept that's
Re: (Score:2)
Re: (Score:2)
Nonsense yourself. They're trained on dumps of the internet.
Re: (Score:2)
I'm not sure what definition you using for interpolation in the high dimensional space in which these algorithm operate. A conventional definition is that nearby points (in the input space) are used to interpolate values in the output space. It's hard to do this in the high dimensional spaces which is why many older ML algorithms fail from the curse of
Re: (Score:2)
Yeah, they learn patterns over the input space, but as soon as you leave the bound of examples in the input space, they will either fail spectacularly, or they will continue the pattern (either appropriately or inappropriately).
Re: (Score:2)
All the interest in deep learning is because they don't fail spectacularly. The patterns they learn generalize the data, but there is no real theory to explain this.
Going back to the XOR example, it's interesting that the network doesn't perfectly generalize. Maybe it gets 99.9
Re: (Score:2)
This class from Stanford is online and explains very clearly how they fail, and why [youtube.com]. Watch it, get educated, and you'll have something more sophisticated to talk about than XOR, which was news from like 40 years ago.
Re: (Score:2)
I'm talking about the generalization performance. They learn patterns that successfully generalize to new data. Yes it's not perfect, but it's learning an effective pattern that covers a large part of the input space/manifold using more than just interpolation.
As for the adversarial examples you cite, these examples are generally created by searching with the gradients. They have essentially no effect on generalization performance because they have close to zero pro
Re: One little problem with their argument (Score:2)
Re: (Score:2)
Maybe you have it backwards. Maybe I've thought deeply about this topic for over thirty years. Maybe I've written papers that have been published in most of the top ML conferences. Maybe I just coauthored a paper on adversarial machine learning. Maybe my job is to do research in these areas.
You're a smart guy, but you post too much. Slow down and think more.
Re: One little problem with their argument (Score:2)
Re: (Score:2)
I never said they could extrapolate, I just said they could do more than interpolation. You're the one who wants to use these low dimensional terms to describe a high dimensional problem. My point is NNs can learn patterns that are beyond what interpolation does. XOR is a simple example that you can code yourself and verify.
Re: (Score:2)
they've never had to deal with a "back end process that consumes protected material, analyzes it and generates new content that competes in the marketplace with the original protected works.
Except about 98% of the artists in that same time span. Creative works seldom spring into existence in a vacuum.
Re: (Score:2, Insightful)
Interestingly, that has actually been a thing since music was invented. It's called students. They listen to music, they learn to analyse that music and to understand that music. They gain appreciation of the styles. When learning an instrument, it is quite common to play other people's music and even to learn to imitate their technique and style. Then they quite possibly become musicians.
If training AI is infringement, then literally every musician out there since the dawn of time is an infringer. We even
Re: (Score:2)
Re: (Score:2)
Re: (Score:3)
Joe listened to some number of hours of other music. The AI harvested [thousands, millions] of hours of music.
Joe produced 1 hour of derivative music using 50 hours of his time. the AI cranked out X hours of derivative music in 40 seconds.
That's the difference.
Re: (Score:3)
I hope you are right about courts recognizing that automated harvesting of "training data" is different from Joe Musician listening to something and producing a similar/derivative work. The SCALE and SPEED that automation brings is the difference.
"Speed and scale" are not relevant: what's relevant is whether the generative AI producing the work can be considered "creative" and the degree of creativity the AI would have put in the generated work.
If the AI is considered to lack creativity, there would be no originality in the generated work, meaning that it would violate the original work's copyright and could not be copyrighted itself.
If the AI is considered to have creativity, the generated work could be considered either derivative or transformativ
Re: (Score:2)
AI techniques have a massive, and unfair, advantage vs. Joe Musician.
Vacuuming up Joe's music is not Fair Use.
Re: (Score:3)
You're missing my point. I would suggest that Speed and Scale is completely relevant, and must be factored into Fair Use now. AI techniques have a massive, and unfair, advantage vs. Joe Musician. Vacuuming up Joe's music is not Fair Use.
Why not? Assuming that the resulting work is original enough to be transformative, it would not matter the speed and scale: the work would be still original and transformative.
Are you arguing producing new, original works should be arbitrarily bottle-necked?
Re: (Score:3)
We have developed easy way to execute a test for detecting copyright infringement in AI databases: 1) remove all content that was not properly licenced from original owner of the material 2) if your product still works, you're ok. If it doesn't work, you're infringing.
I'm not sure how that alone detects infringement (especially, but not limited to, the fact that copyright cases, and what is or isn't infringing is usually determined on a case by case basis, hence why people have confusion over the complexitites of fair use all the time, especially for instance on YouTube).
Re: (Score:1)
Re: (Score:2)
What on Earth are you talking about?
1) Your query was taken down because you were asking about child porn, not because of some sort of evil AI conspiracy.
2) The "Scarlet Johansson case" has nothing to do with AI image scraping - a problem, I should add, where one of the main difficulties is that so many websites strip metadata when you upload an image (and that images are constantly changing dimensions, compression, cropping, adding overlays, etc). Plus the fact that you don't know if the person who posted
Re: (Score:1)
Re: (Score:2)
AI issues are hard, and will not be solved on slashdot.
Re: (Score:1)
The music industry should be like... (Score:2)
Re: (Score:1)
Re: (Score:2)
Re: (Score:2)
I have great fun using it for making memes, such as responding to tankies [youtube.com], responding to a Bluesky dev [youtube.com] user confused about why another user found fish in their yard, responding to a different user [youtube.com] who was confused about how I had a song response to the previous post, turned said dev's text rap battle into an actual rap dis track [youtube.com] (with a bunch of Bluesky inside jokes at the end) and making a dubstep version [youtu.be] of one of his posts about how he's now "DJ Spinster"; turned a comic artist's comic into a song [youtu.be]; turne
Re: (Score:1)
Re: (Score:2)
Re, the rap track: Yeah, I was going for classic early 80s rap. Not sure what you mean by "noise", as it has many meanings. "Recording" static? Compression noise? Record scratching? The breakdown at ~1:50? (was going to remove the latter one but when it's just for a meme it's no worth spending too much time ;) ). If static noise or compression noise, this was before I had a subscription (with a subscription you can download as wav, including stems... though I also can split to stems myself with demucs
Re: (Score:1)
Re: (Score:2)
Hehe, that's the one joke that everyone here knows ;)
It was fun talking to you - I'll keep an eye out for your posts in the future :)
Re: (Score:2)
Congratulations Tsofmia Neptlith, your post is now a song. ;) [youtube.com]
I was going to make it one of your preferred genres, but "Velvet Acid Christ" isn't a genre, it's a band which I can't get myself to enjoy, and Witch House mostly sounds like noise to me so I wouldn't be able to judge whether something is "good" or not. Also sorry if I got your name wrong; I have no clue how it's supposed to be pronounced!
(If I were taking this seriously, I'd inpaint a bit more for cleanup and maybe do a bit of manual post after
Re: (Score:1)
Re: (Score:2)
Spelling fixed. And will check out your recommendations :)
Re: (Score:2)
Can-of-Worms (Score:2)
Re: (Score:3)
They literally filed an amicus brief on this exact topic in the Stairway to Heaven case several years back, arguing that you can only view "derivative" with respect to music in only the most narrow sense or the whole concept of music creation will collapse.