We’re less than six weeks from the end of 2023, and the debate around AI music only seems to be intensifying.
One of the biggest points of contention about the technology is the potential for generative AI, often trained on large unlicensed datasets, to infringe music copyright.
Universal Music Group has been one of the most outspoken advocates for the protection of copyright amid the proliferation of AI-powered voice-mimicking, and melody, chord and lyric-generating technology this year.
Last month, UMG’s publishing division Universal Music Publishing Group sued multi-billion-dollar-backed AI company Anthropic for the alleged “systematic and widespread infringement of their copyrighted song lyrics” via its chatbot Claude.
The suit, filed by UMPG along with co-plaintiffs Concord Music Group and ABKCO, claims that “in the process of building and operating AI models, Anthropic unlawfully copies and disseminates vast amounts of copyrighted works — including the lyrics to myriad musical compositions owned or controlled by Publishers”.
Last week, the three publishers asked a court in Nashville for a preliminary injunction to prevent Anthropic’s AI from using their works while the case proceeds.
A couple of weeks ago we reported on Anthropic’s recent submission to The United States Copyright Office (USCO) in response to a request for written submissions as part of a study around copyright law and policy issues raised by artificial intelligence systems.
That study was announced back in August, when the USCO issued a notice of inquiry (NOI) in the Federal Register on the topic of copyright and AI.
In addition to Anthropic, companies that submitted written responses as part of the study include tech giants like Meta, Google and Adobe, as well as other prominent AI firms like Stability AI.
MBW has spotted that Universal Music Group submitted its own written response to the Copyright Office’s Notice of Inquiry concerning artificial intelligence and copyright.
The 99-page document, which was made publicly available on November 1, contains a series of questions by the USCO and corresponding answers from UMG.
As pointed out in our coverage of Anthropic’s USCO submission, we must stress here that neither set of responses – UMG’s nor Anthropic’s –has anything to do with the lawsuit filed by UMPG against Anthropic.
But the answers within these filings provide an interesting insight into the potential arguments they could use in court.
Here are five observations from UMG’s USCO submission…
1. UMG argues that Training AI using unlicensed songs and recordings ‘is anything but fair use’
As reported previously, in Anthropic’s submission to the USCO in response to its Notice of Inquiry concerning artificial intelligence and copyright, the company argues that using copyrighted material to train LLMs (large language models) is “fair use”.
Anthropic cites various cases, which you can see on page 7 of its USCO filing here, that it argues, “have allowed copying works in order to create tools for searching across those works and to perform statistical analysis”.
UMG does not have the same opinion when it comes to the premise of ‘fair use’ in training AI models.
In a written answer to a question posed by the USCO that asks, “under what circumstances would the unauthorized use of copyrighted works to train AI models constitute fair use?”, UMG is unequivocal in its response:
“The wholesale appropriation of UMG’s enormous catalog of copyright-protected sound recordings and musical compositions to build multibillion commercial enterprises is anything but fair use,” says UMG.
It adds later in the section that, “We can think of no precedent for finding this kind of wholesale, commercial taking that competes directly with the copyrighted works appropriated to be fair use”.
UMG argues that “such appropriation” – training AI systems with copyrighted material without permission – “falls outside of the broad justifications that typically characterize fair uses – such as parodic comment or factual news reporting – and does not satisfy the statutory standards for fair use”.
It adds: “It is simply theft on an unprecedented scale that threatens the core of our business and our mission.”
“The wholesale appropriation of UMG’s enormous catalog of copyright-protected sound recordings and musical compositions to build multibillion commercial enterprises is anything but fair use.”
Universal Music Group
UMG goes on to stress that “the music industry largely operates on a licensing model” and that the company licenses “content for a myriad of purposes, ranging from exploitation of our works for use in particular media (streaming or audiovisual works, for example), to using our content to develop important new technologies, such as AI-powered tools that help make recommendations to users who show preferences for certain kinds of music”.
It adds: “In all these instances, the authorizations we grant support our existing artists and songwriters and provide the resources to discover and develop new artists and songwriters. These licenses, in other words, reward the very creativity copyright law is designed to promote. Allowing AI to swallow our content whole for training without authorization destroys that fundamental purpose.”
Last week, within UMG’s supporting Memorandum submitted to the court when asking for a preliminary injunction during its case against Anthropic, Universal told the court that, “Anthropic will no doubt try to defend its blatant infringement by claiming fair use”.
In that document, which you can read in full here, UMG added that Anthropic “may argue that generative AI companies can facilitate immense value to society and should be excused from complying with copyright law to foster their rapid growth.”
The question of ‘Fair use’ of copyrighted material in AI training was also raised last week when prominent AI exec Ed Newton-Rex explained in an op/ed why he was exiting his role as VP of Audio at Stability AI (which last year raised USD $101 million at a $1 billion valuation.)
As Newton-Rex explained in his op/ed, his personal respect for copyright had clashed somewhat with that of Stability AI in recent weeks. Like Anthropic, Stability AI argued in favor of the ‘fair use’ of copyrighted material to fuel generative AI within its own submission to the US Copyright Office.
2. Universal is ‘adamantly opposed to an opt-out system’ for training AI models using copyrighted content.
In addition to the ‘fair use’ question, the USCO asks if copyright owners should “have to affirmatively consent (opt-in) to the use of their works for training materials, or should they be provided with the means to object (opt-out)”?
The notion of an opt-out program for copyright holders – whereby AI companies can use rightsholders content to train AI models unless those rightsholders explicitly opt out – was reportedly proposed by Google to regulators in Australia earlier this year.
UMG says that it “is adamantly opposed to an opt-out system, where copyright owners are required affirmatively to object to the use of their works for training”.
The company adds: “An opt-out system is based on the erroneous premise that training on copyrighted works without permission is by default lawful unless each copyright owner objects.
“That philosophy does violence to basic principles of copyright law, imposes undue burdens on copyright owners, creates the wrong incentives for AI developers, and is neither practicable nor effective for protecting the rights of copyright owners or ensuring the sensible use of copyrighted works for training purposes.”
“An opt-out system is based on the erroneous premise that training on copyrighted works without permission is by default lawful unless each copyright owner objects.”
Universal also offers a bit of copyright 101, citing Section 106 of the Copyright Act which “grants copyright owners the exclusive right ‘to do and to authorize,’ inter alia, the reproduction, distribution, and performance of their copyrighted works.”
UMG continues:: “To ‘authorize’ means ‘to give official permission for something to happen, or to give someone official permission to do something’.
“The only way to grant ‘official permission’ is for a copyright owner to express affirmative consent before a third party exploits the owner’s copyrighted work’. Copyright law works on this basis, and silence, much less unawareness of an exploitation, cannot substitute for knowing consent.”
Universal argues in its written response to the USCO that “there is no equitable reason why AI companies merit exemption from this baseline requirement of copyright law. Moreover, an opt-out system would place an insuperable burden on copyright owners.”
3. “UMG is ready and willing to explore licensing content to generative AI companies”
One of the questions in the USCO’s questionnaire is whether or not “direct voluntary licensing [is] feasible in some or all creative sectors”.
In other words, is it “feasible” for an AI company to pursue licenses directly from companies whose copyrighted content they wish to use to train their AI models?
Universal’s answer is that “the licensing of music for training generative AI is not an insurmountable administrative task or an undue burden on the AI community”.
The company adds: “Any contrary belief is grossly mistaken. Licensing is the primary model by which the music industry does business, in connection with both traditional markets and the newest technologies.”
“The licensing of music for training generative AI is not an insurmountable administrative task or an undue burden on the AI community.”
UMG claims to be a “leader in establishing business models for cutting-edge technologies,” which it says it “confronts on a regular basis” and points to streaming music and social media as two examples
“Despite this seismic takeover of music consumption by previously unknown technologies, the entire music industry managed in short order to license virtually all of its content to all of the major streaming music services, including Apple Music, Amazon Music Unlimited, Deezer, SoundCloud, Spotify, Tidal, TikTok, YouTube, and many others,” says UMG.
“In fact, Universal Music Group has music licensing arrangements with hundreds of different digital music businesses worldwide.”
UMG adds that it is “ready and willing to explore licensing content to generative AI companies and has already established our disposition to work with those companies to find ways to build fair, lawful, and ethical business models that make economic, strategic, and long-term sense for our artists and songwriters.”
A recent example of Universal partnering with a generative AI company arrived in May when UMG teamed up with AI startup Endel to create “AI-powered, artist-driven functional music”.
4. Universal “does not believe that a human being using a generative AI system can qualify as the author of outputs of that system”.
UMG says that it “does not believe that a human being using a generative AI system can qualify as the author of outputs of that system, because the output is never copyrightable”.
Thar response came in answer to this question from the USCO: “Under copyright law, are there circumstances when a human using a generative AI system should be considered the “author” of material produced by the system?”
The company continues: “Authorship requires human expression, and the output of generative AI systems does not embody human expression. As noted in the NOI at 1 n.1, generative AI systems produce outputs based on learning statistical patterns in existing data.
“When generative AI responds to a human text prompt, it draws upon its analysis of those statistical patterns to create something that matches or ‘autocompletes’ the human prompt, but it is the AI system that creates the output, not a human being.
“Whatever the nature of the output – text, music, image – its expressive content was created and arranged by a computer making statistical predictions rather than creative choices.
“In this regard, it is important to distinguish between the potential copyrightability of a prompt and the output it might generate. The former might be copyrightable, but the latter is not.”
Elsewhere in this section, UMG explains that “many prompts may fail the test for copyrightability because they amount to no more than a non-protectible idea”.
It adds: “Instructing a generative AI system to create a ‘poem about pine trees”’ or an image of a ‘lion playing an electric bass’ provides only noncopyrightable concepts.
“However, a prompt may cross the threshold of copyrightability to the extent it satisfies the minimal standards of originality and creativity set forth in Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991).”
5. UMg believes there should be “a federal statute” to “prohibit unauthorized generative AI voice clone models”
Last but not least, one of the responses that stood out from UMG’s submission to the USCO was about federal and state laws around the rights of publicity.
The USCO asks if Congress should “establish a new federal right, similar to state law rights of publicity, that would apply to AI-generated material?” UMG appears, from its response, to be very keen for such a law to be introduced.
“Yes, Congress should enact a federal right of publicity statute that would apply to content generated by AI,” notes UMG in its submission.
“Rather than preempt state right of publicity laws, a federal statute should establish a floor for the protection of an individual’s name, image, likeness, and voice.”
Universal explains that in instances “when AI has been used to steal or appropriate an artist’s persona,” the company has “asked distribution platforms hosting, displaying, and transmitting that content to remove the material on the grounds that it violates state right of publicity laws”.
To date, according to UMG, “the response to such requests has been uneven, and some services have refused to remove the material citing claims of Section 230 immunity.”
“A federal statute should prohibit unauthorized generative AI voice clone models.”
In addition to a federal right of publicity statute applying to content generated by AI, UMG argues that “a federal statute should prohibit unauthorized generative AI voice clone ‘models’.”
In other words, “algorithms designed with the intent of allowing users to create new recordings in a mimicked voice (or voices) used to train the model without consent.”
This particular suggestion arrives a few months after a media storm around Universal Music Group-affiliated megastars Drake and The Weeknd seeing their vocals replicated by AI tools without permission, and then ‘performed’ within a viral track called heart on my sleeve, created by the ‘artist’ called ghostwriter.
Via a third-party distributor, the track made its way to streaming services like Spotify, Apple Music, SoundCloud, Deezer and TIDAL, where it began to rack up streams by the hundreds of thousands. It was subsequently deleted by these platforms.
UMG adds in its recent USCO submission: “A meaningful statute will also include a First Amendment balancing test to harmonize the potentially conflicting interests of the individual and free speech.
“We firmly believe in freedom of expression, but broad categorical exemptions in some existing state statutes can enable the exception to swallow the rule.”
It continues: “A court is in the best position to weigh the facts and balance the interests; defendants should not be able to trample on an individual’s right of publicity by simply labeling their AI creation a form of content included in an all-encompassing list of categorical exemptions.
“Beyond those contours, it is important that a federal statute (1) institute fines and enable damage recoveries sufficient to establish a deterrent effect, (2) authorize both federal enforcement and a private right of action, and (3) as with all forms of intellectual property, the right should be eligible for assignment or licensing either in whole or in part, so that enforcement may be delegated.”
