They should start with AudioCraft itself, conceptually it's derivative work and it doesn't matter if it's "open source" or not. Try throwing in someone's sample in a song and publish it saying "no copyright infringement intended and I totally don't make any money from it"... If it becomes popular, see how long it stays up until DMCA takedown. And we know this dataset is already popular.
Even before streaming, You never "owned" any music legally [1], you merely owned a physical copy of a performance[3] of a song, that in no way gives you the right to make derivative works [2] automatically.
Also it doesn't really matter on what the law says, RIAA in the last iteration, relied on the fact it you would rather pay a fine than be able to afford expensive lawyers to fight the specifics out in court on average.
It was always about disproportionate ability to bring resources against individual "offenders" to create fear among everyone to deter "undesirable" forms of copying, not necessarily what the legal protections were.
---
[1] Unless you specifically commissioned it under a contract which gave you the right
[2] See recent cases including those related to Kris Kashtanova and Andy Warhol.
> that in no way gives you the right to make derivative works
True, but only because you have that right anyway. I can do anything I like with copyrighted content I legally possess, as long as I don't distribute the results of my efforts.
Establishing derivation is at the crux of all legal matters surrounding diffusion models. It has not yet been clearly established. If it is, then I'd agree with you. Until then, I think it's a bit more up in the air.
Also, IIRC, RIAA did not bring many resources to bear against e.g. "home taping" itself, because they could essentially never know that it had occured. The overwhelming majority of their efforts went into trying to takedown people distributing multiple copies.
The Kashtanova case does not cover derivation in any real way, but is really about copyright attribution choices between human and software.
The Warhol case specifically tests a fair use claim, not a derivation claim.
RIAA and others in general sued a lot of people including bar owners for playing their songs etc , there were also some enforcement via private companies with three strike policies and so on especially in Europe .
They went after ISPs and torrent sites which only hosted magnet links and many others who shouldn’t have been really sued
The goal was to create a very hostile environment for downloading songs to protect their interests - “you wouldn’t download a car!”
It was never the goal nor ever realistic to actually pursue enforcement action against every offender , the idea was to change behavior with all the related actions .
They did end up changing behavior, people just didn’t want the hassle, or be in fear so paying for streaming for access had a stronger value proposition, it was not what the RIAA planned , but benefits enormously today.
They may have gotten the AHRA passed, but they essentially lost in every important way:
> "This exception was crucial in RIAA v. Diamond Multimedia Systems, Inc.,[14] the only case in which the AHRA's provisions have been examined by the federal courts. The RIAA filed suit to enjoin the manufacture and distribution of the Rio PMP300, one of the first portable MP3 players, because it did not include the SCMS copy protection required by the act, and Diamond did not intend to pay royalties. The 9th Circuit, affirming the earlier District Court ruling in favor of Diamond Multimedia,[15] ruled that the "digital music recording" for the purposes of the act was not intended to include songs fixed on computer hard drives. The court also held that the Rio was not a digital audio recording device for the purposes of the AHRA, because 1) the Rio reproduced files from computer hard drives, which were specifically exempted from the SCMS and Royalty payments under the act, 2) could not directly record from the radio or other transmissions. "
From the AHRA itself:
> No action may be brought under this title alleging infringement of copyright based on the manufacture, importation, or distribution of a digital audio recording device, a digital audio recording medium, an analog recording device, or an analog recording medium, or based on the noncommercial use by a consumer of such a device or medium for making digital musical recordings or analog musical recordings.
and from Wikipedia again:
> In regard to home taping, the provision broadly permits noncommercial, private recording to analog devices and media. However, it fails to resolve the home taping debate "conclusively," as it only permits noncommercial, private recording to digital devices and media when certain technology is used.
> Two reports by the House of Representatives characterize the provision as legalizing digital home copying to the same degree as analog. One states "in the case of home taping, the exemption protects all noncommercial copying by consumers of digital and analog recordings,"[22] and the other states "In short, the reported legislation [Section 1008] would clearly establish that consumers cannot be sued for making analog or digital audio copies for private noncommercial use."[23]
Similarly, language in the RIAA v. Diamond Multimedia decision suggests a broader reading of the Section 1008 exemptions, providing blanket protection for "all noncommercial copying by consumers of digital and analog musical recordings" and equating the spaceshifting of audio with the fair use protections afforded home video recordings in Sony v. Universal Studios:
>> In fact, the Rio's operation is entirely consistent with the Act's main purpose – the facilitation of personal use.
Legal acquisition does not matter for AI training. If training is fair use then you can train on pirated material (e.g. OpenAI GPT). If it's not fair use then buying the material does not matter, you have to negotiate a specific license for AI training for each work in the training set, which is impractical at the scales most AI companies want to work.
If you purchase the music, you have a (sometimes explicit, sometimes implicit) license to do certain things with the music, entirely independent of any concept of "fair use".
The question is not "is training part of fair use?" but "is training part of, implicitly or explicitly, the rights I already have after purchase?"
Given that "training" can be done by simple playing the music in the presence of a computer with its microphone turned on, it's not clear how this plays out legally.
In the US, exceptions to copyright come across in two distinct bundles: first sale and fair use. They exist specifically because of the intersection between copyright law and two other principles of the US constitution:
- First sale: The Takings Clause prohibits government theft of private property without compensation. Because copyright owners are using a government-granted monopoly to enforce their rights, we have to bound those rights to avoid copyright owners being able to just come and take copies of books or music you've lawfully purchased.
- Fair use: The 1st Amendment prohibits government prohibitions on free speech. Because copyright owners are using a government-granted monopoly to enforce their rights, we have to bound those rights to avoid copyright owners being able to censor you.
If you hinge your argument on "I bought a copy", you're making a first sale argument.
Notably, first sale is limited to acts that do not create copies. This limit was established by the ReDigi case[0]. Copyright doesn't care about the total number of copies in circulation, it cares about the right to create more. So an AI training defense based on first sale grounds would fail because training unequivocally creates copies.
Fair use, on the contrary, does not care if you bought a copy of a work legally. It only cares about balancing your right to speech against the owners' right to a monopoly over theirs. And it has so far been far more resistant to creative industry attempts to limit exceptions to copyright - to the point where I would argue that "fair use" is an effective shorthand for any exception to copyright, including ones in countries that have no fair use doctrine and do not respect judicial precedent.
The courts won't care how the training comes about, just if the act of training an AI alone[1] would compete with licensing the images used in the training set data.
Is training a "pirate model" something you'd reasonably be able to do at home though, given the compute requirements? The analogous "image generation at home" is only possible due to a for-profit entity with significant resources choosing to (a) play fast-and-loose with the provenance of their training set and (b) giving away the resulting model for free, if the open source community had to train their models from scratch then as best as I can tell they would still be stuck in the dark ages generating vague goopy abominations.
Currently, yes, available compute power @ home does indeed seem like a limitation. Whether that remains true going forward seems a little unclear to me.
Ugh, I dread having to listen to everyone's hyper personal music because they swear up and down to the point of tears that "IT'S THE BEST SONG EVER CREATED! EVER!!!", while the constantly prod for you to affirm how amazing the song is.
Bruh, music is subjective as hell, and I can already tell I hate this song.
Perhaps LoRA (Low-Rank Adaptation) training techniques could be used for these types of models, like they're currently being used with LLMs and latent text-to-image diffusion models.
Sadly looks unlikely if the base model wasn't trained on vocals.
> Mitigations: Vocals have been removed from the data source using corresponding tags, and then using a state-of-the-art music source separation method, namely using the open source Hybrid Transformer for Music Source Separation (HT-Demucs).
> Limitations: The model is not able to generate realistic vocals.
Something like the algorithm TikTok uses. First probing by offering a variety of content that should match based on what little information you have on the user (ip location, locale, etc).
Then use the user’s action to iteratively refine your classification, until you end up with something tailor-made.
That'll make musicians happy with big tech as well, just like artists are. *sigh*