Copyright & Artificial Intelligence Listening Session - Music & Sound Recordings
>> Andrew Foglia: Hello, everyone, and thank you for joining us today for the Copyright Office's listening session on AI and Music and Sound Recordings. I'm Andrew Foglia, Deputy Director of Policy and International Affairs. To kick off today's listening session, it is my pleasure to introduce Shira Perlmutter, Registrar of Copyrights and Director of the U.S. Copyright Office. Shira? >> Shira Perlmutter: Thank you, Andrew.
Welcome to the Copyright Office's public listening session on artificial intelligence, this one focusing on musical works and sound recordings. This is the fourth and final of this series of listening sessions on AI and copyright. While some of today's discussion may overlap with prior listening sessions, we recognize that the music community has unique perspectives, concerns, and use cases.
The Office appreciates the high level of public engagement with these listening sessions. We've received hundreds of requests to participate, and thousands of people have watched remotely. This interest is, of course, a reflection of the astonishing potential of artificial intelligence and the impact that it's already having in our lives and on society as a whole. So to start off, I can provide a few takeaways from our prior listening sessions. First, there is disagreement about whether or under what circumstances training generative AI on copyrighted works could be considered fair use. Now, of course, since the last listening session, the Supreme Court has issued a new fair use decision in Andy Warhol Foundation v. Goldsmith,
which will have to be taken into account in such discussions going forward. Second, there's considerable interest in developing methods to enhance transparency and education with respect to how generative AI produces works, including the possibility of tracking relationships between ingested works and outputs, and also understanding how assistive AI is used as a tool in the creation process. And finally, many stakeholders have questions about the Office's registration guidance for works containing AI-generated material and would like more details and more examples of how the Office will approach applications for such works. On this last point, the Office will host a public webinar on June 28th, where our registration experts will walk viewers through our registration guidance and answer frequently asked questions. We'll then host a second webinar on July 26th to focus on international perspectives on AI and copyright. And as we've mentioned before, the Office will be issuing a notice of inquiry in the coming months, seeking public comments on many of the issues raised during these listening sessions.
Now, today's session focuses on musical works and sound recordings, and certainly the music industry has a long history of employing the latest technologies to create new works, from distortion pedals, to digital audio workstations, to spatial audio, and even autotune. So there's a lot of interest in how the use of generative AI is similar and how it is different. Before I hand over the virtual mic, let me thank our panelists in advance for contributing to today's conversation.
This is a complex and very important topic, and one that has great personal significance for many of our participants. Your perspectives are critical in informing sound public policy, and we look forward to an enlightening discussion. I will now turn things back to Andrew Foglia for more information about today's session. >> Andrew Foglia: Thank you, Shira. So as Shira mentioned, today's listening session is the fourth and final in the Copyright Office's series of AI listening sessions. These listening sessions will inform further steps in the Office's AI initiative.
Questions our panelists raise may be ones in which we seek written comments later this year. For further information about our initiative, to learn about upcoming events or watch past listening sessions, please visit copyright.gov/ai. So today's session will consist of two panels with a brief break in between, and a few Zoom housekeeping points before we begin. First, if you are joining this session but are not a panelist for this particular session, please keep your camera turned off and your mic on mute.
Second, we are recording this session today. The recording will be available about three weeks after today's session. Third, the transcription function is activated as well. So today's panels will start with a brief introduction and a short statement by each panelist.
We request that these statements be limited to three minutes. Moderators will be watching the time, and if you go over your allotted time, we will have to cut you off to reserve time for other participants. After these introductions, we will have a moderated listening session. The moderator questions, most of which the panelists have received in advance, are intended only as prompts for discussion, and we welcome participants to share any relevant perspectives and experiences that they feel are important for the Office to hear. Panelists who wish to speak should use Zoom's raise hand function, and our moderators will try to call on you in the order in which you raise your hand.
I do want to emphasize that this is a listening session and not a debate. There will be other opportunities for participants to engage more directly with competing views and questions from others. Finally, we will not be accepting questions from the audience. If you are in the audience and you want to share a question or comment with the Copyright Office, we will be soliciting written comments through a notice of inquiry later this year. With that, I will hand it over to our moderators for the first session. Jason Sloan is an Assistant General Counsel in our Office of the General Counsel.
Chris Weston is a Senior Counsel in our Office of Policy and International Affairs. The mic is yours, Jason. >> Jason Sloan: Thanks, Andrew. Welcome, everyone. We'll begin with introductory statements in the order stated on the agenda.
As we asked you in advance, as part of your three-minute statement, please tell us what you think is most important for us to know about the use of generative AI in the music industry. For example, how is it being used? What are the opportunities and challenges, advantages, or disadvantages? And what do you foresee to be the near and long-term industry impacts? Let's start with Nathaniel. >> Nathaniel Bach: Hi. Good morning. Good afternoon.
I'm Nat Bach, an entertainment litigation partner at Manette Phelps in Los Angeles. I'm here today representing Music Artists Coalition, or MAC, an organization dedicated to putting artists' rights first. Our membership includes a diverse roster of both contemporary artists and icons, like Don Henley, Dave Matthews, Anderson pack, Billie Eilish, Diplo, Bonnie Raitt, and Neil Young, to name just a few. Generative AI presents a profound opportunity and challenge for artists in music.
On the one hand, musicians have always embraced new technology, but on the other, those advances have been in service of creativity that starts with a human hand and ear. I'd like to make six brief observations in my opening remarks. First, human artistry should prevail over machine-based shortcuts every time. The technology we are facing today will change and evolve, but by asking ourselves at key junctures how we can protect human artistic creation and support artists, we can remain on the right path. Second, music is different. It is different than visual arts in its ability to elicit emotions, and the power of a song to tell a story, on its own or as part of a film, dramatic work, or television commercial, is unparalleled.
Music is also different as it relates to how AI models are and can be trained. Unlike the billions of images on which some text-to-image AI platforms are trained, including vast numbers of images in the public domain, the universe of recorded music is smaller and generally accessed via portals and DSPs like YouTube, Spotify, Apple Music, and others. Those seeking to train AI models and scrape songs are likely to do so off of these types of services, which also play an important gatekeeping role. Third, the training of AI models on artists' works without a license is infringing and not a fair use. The Supreme Court's decision in Warhol Foundation v. Goldsmith confirms the primacy of artists' songs and recordings where the purpose of the use is similar, and under its reasoning, AI tools that scrape, ingest, or copy such musical works are not transformative.
Gen-AI companies may use language to suggest transformativeness, but fulsome disclosures will be needed to determine what is actually happening under the hood. Fourth, primarily AI-generated music threatens the already meager royalties that artists can earn via streaming. The larger the slice of the streaming pie that is taken up with AI-created functional music, the less royalties that DSPs will be required to pay out to human artists and their affiliated licensees. Human artists should be incentivized to create music, but diluting the royalty pools in such fashion would have a significant negative impact.
Fifth, copyright is a key pillar of artists' rights and protections, and should work in tandem and not at cross-purposes with other artists' rights, like the right to publicity and rights under the Lanham Act. The Act supports enactment of a strong federal right of publicity law that will protect persona and identity, but that does not lessen the need for robust copyright protection. Sixth, and finally, we cannot be blinded by the allure of technological advancement without thinking through its ramifications. In the past, lawmakers have failed to protect copyright because they were seduced by changing technology. We should not be misled or confused by those who claim that AI itself is about freedom and creativity, and not in service of human creators, especially where profit motives threaten the artists and songwriters on whose backs they may build their businesses. Let's not be fooled again.
>> Jason Sloan: Thanks, Nat. >> Nathaniel Bach: Thank you very much. I look forward to the conversation.
>> Jason Sloan: Thank you. Timothy? >> Timothy Cohan: Thank you. My name is Tim Cohan. I'm the Chief Counsel for Pure Music, a global international music publisher.
I'd like to thank the Copyright Office for its timely focus on artificial intelligence and for the opportunity to share the perspective of one publisher and the songwriters we represent. When we ask our writers whether they're using AI, they're not sure. Compositional tools that employ AI are already integrated into the creative process. Whether these tools today generate traditional elements of authorship is unclear.
What is certain from our point of view is that all of our writers will be using AI in some form before long. We're grateful to the Copyright Office for raising important issues around the use of AI in the creative industries in its March 2023 guidance, which provoked remarkable engagement and discussion in the music community. That said, on close review, we have some concerns about the potential impact on the protection and promotion of creative works if the current guidance were implemented to the letter.
As we read the guidance, we must require our writers to specifically identify any AI in delivered works. Some publishers will simply refuse to accept such works. Either way, we risk creating a tripwire to breach of contract that doesn't exist today.
A writer may believe the use of AI was de minimis in good faith and fail to disclose, or may be entirely unaware of the use of AI in a song by a co-writer or producer, for example. Moreover, we're not optimistic about getting precise AI metadata when we still struggle to receive songwriter splits, not to mention the essential song metadata that the marketplace desperately needs. The Copyright Office has suggested that in cases of uncertainty, we may register a work with a general statement that it contains AI-generated material. This is helpful.
However, under current guidance, we then need to wait for the Copyright Office to contact us to discuss each such work. Unless we're prepared to inspect and analyze every song delivered with AI content, this would appear the only feasible option. As a publisher, we protect our writers' works through timely registrations with the Copyright Office. Works not registered do not receive the full statutory protections of the Copyright Act. Thanks to the modernization efforts of the Copyright Office, this critical process has become more streamlined and efficient. We would hope not to move in the opposite direction with a process that can't be scaled and yet may well become the norm.
We would respectfully urge policymakers to afford writers the presumption of authorship in the selection and arrangement of AI material in works submitted for copyright protection. If a question arises as to the enforceability of a particular copyright, then, if necessary, a court can conduct a relevant fact-specific inquiry. If that inquiry must instead take place through patent-level scrutiny of every registration at inception, the result may be an effective prohibition on the use of AI in the creative process. We know that prohibitions on the use of technology have not historically turned out well.
Putting the question simply, do we want to foster the creative process or fence it in? We would err on the side of creativity and the presumption of protection. Thank you. >> Jason Sloan: Thanks, Tim.
Kenneth? >> Kenneth Doroshow: Good afternoon. My name is Ken Dorsho. I am the Chief Legal Officer of the Recording Industry Association of America. The RIAA's record company members create, manufacture, and distribute sound recordings representing the majority of all lawful consumption of recorded music in the United States, including many of the most popular and commercially valuable sound recordings in the world.
The recording industry is and has always been a tech-forward business, ever since the invention of the phonograph through the eras of vinyl, tapes, CDs, and now streaming. Record companies appreciate the valuable role that new technologies, including artificial intelligence, can play in the creative process. Indeed, AI and machine learning is already in use in many facets of music production and distribution. Like every new technology, AI will undoubtedly push creative boundaries and help shape recording artists' visions and expand their commercial reach.
We embrace AI's potential as a tool to support human creativity, but not to supplant it. Like any new technology, AI must be used responsibly, ethically, and in a manner consistent with the law. By and large, unfortunately, this is not happening today. Many AI developers, including some of the largest companies in this fast-growing industry, have scraped online music repositories and copied vast troves of copyrighted sound recordings to build datasets for their AI models, and did so, and continue to do so, without the consent of artists or rightsholders. This, in short, is copyright infringement on a massive scale.
And the arguments of fair use that we frequently hear as justification for these practices are misplaced, especially in the wake of the Supreme Court's recent Warhol decision, which Shira mentioned in her opening remarks, and I'm sure we'll discuss in greater detail today. We've also seen a proliferation of AI models and services specifically designed to enable the generation of recordings containing digital replicas of well-known artists' voices and styles. In most cases, this is done without the artist's permission to capitalize on the commercial value of the artist's voice, resulting in a distortion of the artist's own vision for themselves, confusion in the marketplace over whether the artist has endorsed this use of their voice, and a dilution of the value of the artist's brand. RIAA's members believe that free market licensing is the right path forward for the use of copyrighted sound recordings by developers of AI systems.
We already know that free market licensing of sound recordings works. In fact, it's the foundation of today's thriving streaming economy in which all of the leading services have managed to obtain licenses from our member companies. The necessary licensing markets already exist. They are practical, efficient and have a track record of success. There is no reason why they cannot work in the new world of AI.
I'm honored to be included in this listening session. I'm grateful to the office for allowing me to participate in today's very important discussion. >> Jason Sloan: Thank you, Ken.
Jack. >> Jack Kugell: Hey, everybody. I'm Jack Kugell. I'm a Grammy and Emmy-nominated songwriter and producer and a co-founding board member of Songwriters of North America, where I co-chair the advocacy committee. Songwriters of North America is a non-profit membership organization founded in 2015 by a group of songwriters, composers, and music industry professionals determined to advocate for ourselves in an increasingly challenging digital economy.
SONA has since evolved into a trade association and hub for thousands of engaged working music creators representing the boots on the ground songwriters and composers who call making music their job. I'm honored to be a part of today's panel representing SONA. What is most important for the Copyright Office to know about the use of generative AI in the music industry? It is of utmost importance to protect human creators and their rights in the face of developing AI technology. Policymakers must ensure that AI development be done responsibly and in a way that does not threaten the livelihood of human creators, particularly songwriters. Creators must have a say over whether they want their works to be used for AI training. Generative AI often illegally takes copyrighted music written by human creators without permission or compensation.
In doing so, it also purposely removes the work's metadata so it can't be tracked. As this technology develops, it could forcibly generate music that competes in the marketplace with the very human-created music ingested in the first place. Developers will claim that this is a fair use issue. However, fair use was not created to allow the replacement of, nor to compete with, human-created work in the marketplace. We need the Copyright Office and the courts to recognize that our works must be licensed.
Songwriters need to have the choice of whether they want their works to be used by AI companies and the ability and right to say yes or no. How will we know if AI uses our work? We need to have records kept. We need complete record-keeping of what is in the database as well as tracking specific end-user queries, i.e. retinal song-like prints.
Again, this illustrates the need to retain the ingested work's metadata. Human creators should be able to use AI as a tool, as we have done with many technological developments in music in the past, and have the assurance that our works will be protected by copyright. Thank you. >> Jason Sloan: Thanks, Jack.
Garrett? >> Garrett Levin: Thanks, Jason. Thank you to Registrar Perlmutter and the entire U.S. Copyright Office team for inviting me to speak at today's listening session and for your steadfast engagement on this important issue. My name is Garrett Levin. I'm the President and CEO of the Digital Media Association, DMA, the trade association that represents the world's leading audio streaming services. Music has long been at the forefront of potentially disruptive and new technology, and the development of new technology has often been met with initial concern by many in the industry.
However, the success of today's streaming-driven music industry is definitive proof that music and technology can, should, and most often do learn to work together and enrich our musical traditions. Emerging technologies have historically improved the creation, distribution, and consumption of music. AI is a rapidly evolving technology with similar abilities to assist creators, including human musicians and songwriters, and improve the way music is created, distributed, and consumed. But AI is not one-size-fits-all. Current discussions around AI often lack grounding definitions, including the lines between generative and assistive AI.
The entire industry will benefit from establishing a common set of facts in these discussions and are focusing questions around specific technologies and uses. Similarly, policymakers benefit from shared substantive expertise about AI technologies, evolving trends, and the potential effects on artistic expression, innovation, and commercial markets before proposing changes. We hope that the Office's series of listening sessions reflects the start of that kind of analysis, and DEMA members are willing to assist the government in pursuing that evidence-based path. We'll no doubt dig further into some of the specifics during the discussion today, but at a high level, DMA members believe the following. Existing U.S. copyright laws, including those governing copyrightability, such as originality, de minimis contribution, sona fare, and the idea-expression dichotomy, infringement, including questions of unlawful appropriation, substantial similarity, and causation, and the DMCA, as well as laws that exist outside of copyright to protect one's name, likeness, and the right of publicity, are sufficient to address creations made with or by AI technology.
Different legal doctrines can and should be employed to consider the various questions arising from AI-generated music, but copyright law should not be stretched or changed to address questions that more properly arise under laws relating to trademark, right of publicity, or unfair competition. And one final note in this introductory statement on the topic of data, a topic on which DMA members have extensive experience. Music streaming services should not be, and cannot be, the arbiters as to what is or is not AI-generated. It is not possible with the existing data, and any new data must come from copyright owners and creators. Data accountability must exist throughout the entire chain from creation to distribution. There have long been data challenges in the music space.
Tim actually acknowledged these in his opening remarks, including with ensuring that accurate and complete metadata identifiers are included in recordings at the time of distribution. These challenges existed before streaming, continue to exist, and are highly relevant to discussions of the treatment of AI-generated music in streaming. Thank you, and I look forward to the discussion. >> Jason Sloan: Thank you, Garrett. Kevin. >> Kevin Madigan: Thanks, Jason, and thanks to the Copyright Office for hosting these listening sessions and allowing me to participate.
I'm Kevin Madigan with the Copyright Alliance, and I want to make just a few points in my opening remarks. The first of which is that as we consider questions surrounding copyright infringement and generative AI, it's really important that we separate the discussion of potentially infringing output from infringement that occurs when works are ingested by AI systems without authorization. There have been comments made in some past listening sessions about how infringement isn't really an issue because output of generative AI is so rarely substantially similar to the works that are ingested. But even if that's true, it doesn't change the fact that there are unauthorized reproductions occurring at the input or ingestion stage, and the right of reproduction is a standalone right that's implicated.
The second point I'll make is that, especially now in the wake of the Supreme Court's Warhol v. Goldsmith decision, it's essential that we recognize that the purpose of many generative AI systems is to use expressive works of authorship to generate new works. In many cases, that means their purpose is to create works that act as a substitute for the works they ingest. Now, what Warhol v. Goldsmith also makes clear is
that transformative use does not control a fair use analysis. And so claims by some AI developers that the transformative nature of AI means that it just automatically qualifies as fair use are clearly not supported by the law. The last point I'll make is that in earlier listening sessions, some argued that there are cases that support the position that AI ingestion of copyrighted works qualifies as fair use, particularly SEGA v. Accolade and the Google Books case. I disagree strongly that either of those cases would control an AI fair use analysis because they're clearly distinguishable. As we know, fair use is a very fact-specific analysis, and in the SEGA case, which was a reverse engineering case, the court was clear that its analysis was specific to the functional computer-coded issue, whereas AI systems make use of clearly expressive works of authorship.
And then in Google Books, the purpose of the scanning was to provide information about the books, not to create new substitutions for the underlying works. So while these cases may be instructive in some ways, they deal in very different fact patterns that are clearly distinguishable from AI ingestion. So I'll stop there for now, and I look forward to the rest of our discussion. >> Jason Sloan: Thank you, Kevin. Alexander. >> Alexander Mitchell: Hey, Jason.
Thanks so much. I'm Alex Mitchell. I'm a musician. I am a policy advocate for the general music field, and I'm the co-founder and CEO of a company called Boomy.
I want to thank the Copyright Office right off the bat for allowing me to participate today and for being open to a variety of viewpoints on these very nuanced issues. Boomy is a free online platform where creators and enthusiasts all over the world are making, sharing, and monetizing generative music every day. More than a million Boomy creators have already produced over 15 million original songs using our proprietary technology, and a small percentage of those have been released by our creators through Boomy as a record label and a publisher. We founded Boomy on the principle that every human on the planet should be able to express themselves with music, regardless of their access to resources like expensive studio time, instruments, or even high-end computers. And in pursuit of that human expression, we determined very early on to respect copyright as part of our commitment to ethical AI. So what does ethical AI mean? What it means to us is not creating so-called black box models that are trained using third-party data and are going to be widely discussed today.
Instead, we've developed an original generative music framework that is directly inspired, designed, and influenced by the musicians who work at Boomy and design our algorithms. Generative music, to me, represents a new creative class of technology-enabled musicians. And this creates an on-ramp for musical expression that we believe will increase the overall interest and participation in the music industry. This is the big opportunity that I think can get lost in some of these conversations. And as a label ourselves, of course, it's our position that the original songs that our creators are making with Boomy should be subject to the same copyright protection as the songs made by generations of algorithmic musicians or generative musicians, whatever term you want to use, that came before us. These protections aren't just important from a business perspective.
They're crucial for the prevention of harmful content and the mitigation of unauthorized uses of platforms like ours. So these questions of copyrightability of AI music which now has a definition that changes almost weekly are mission-critical for us in supporting a rapidly growing community. My hope is that today and in the future that I can be a helpful voice in this conversation from the perspective of a platform that is very much on the front lines of the incredible opportunities and the difficult realities that come with this new generation of technology-empowered human creative expression. >> Jason Sloan: Thanks, Alex. Rohan. >> Rohan Paul: Hey, everyone.
Thanks for having me. I'm Rohan Paul, an artist and founder of Controlla. Controlla is a platform powered by human singers that helps anyone create, protect, and monetize their AI voice. We're helping major labels and artists pinpoint unauthorized uses of their AI voices on social media to protect some of the world's most beloved singers. Three months ago, I was engaging with various music tech companies to create a platform where listeners could hear any song in any artist's voice.
I believe this platform needed to be done on an opt-in basis where artists would contribute their songs and voices so that fans can generate AI covers with permission. Instead, they started posting AI covers but never got permission from artists. They garnered millions of views on social media and started offering direct access to celebrity voices in their apps without getting permission from a single artist.
I felt so ashamed. The teams that I trusted in hopes of supporting artists turned around and straight-up robbed them. It was at this time various Discords were created and a community of AI creators were training models on celebrity voices making original songs like Heart On My Sleeve. This is no longer bringing attention to the original artists.
It's appropriating their voice and brand for clout. After engaging with people in this Discord, it was clear that some of them knew this was wrong but others didn't believe it was illegal or unethical in any way. They saw it as a form of admiration towards the original artists. They felt like it was their only chance of collaborating with an artist they love that would otherwise never work with them. Many times, they would point to these existing apps and say, and say, if there's an app for it, it must be legal, right? Meanwhile, dozens of companies and apps started stealing these voices and using this new community to kickstart their products and offer easy-to-use celebrity singing voices without permission.
They call them user-uploaded voices but it's obvious that almost every single voice is stolen. These companies didn't create anything. They didn't create the tech, they didn't create the music, and they didn't create the voices that artists spent their entire lives training.
It makes no sense why a platform like this should continue to exist and profit off the backs of artists and researchers simply because they lack the decency to ask for permission. I believe that every single one of these platforms needs to have all unauthorized voice models removed immediately as they already contribute to millions of AI songs each week. Despite these voice-stealing platforms, I do believe AI voices and other AI music tools have tons of value to offer artists, fans, and all players in the industry.
We just need more clarity on what isn't allowed and how artists who want to embrace it should be compensated from derivative works in their voice. Artists should have control of their own voices, and those that want to embrace opening up access to fans should be able to do so on their own terms. My proposed solution would be to explicitly include protection of someone's voice, whether real or AI, under copyright law. Furthermore, I'd advocate that any past contracts that give catalog owners permission to use recordings in any way should not extend to the cloning of a voice or style with AI. We're at the point where AI voices are indistinguishable from the real thing, even by AI.
So we can't compare this to sampling or training systems on other forms of media like text and images, because our voice is as unique as a fingerprint and it should be treated as part of our identity. Thanks. >> Jason Sloan: Thank you. Jason. >> Jason Rys: Hi, my name is Jason Rys.
I'm the co-owner, CTO, and EVP of Wixen Music Publishing. We're a music publishing administration company that represents many of the finest songwriters and songs of the last hundred years from rock bands like Tom Petty, The Doors, Weezer, Rage Against the Machine, to hip-hop icons like Missy Elliot and even to old standards like Santa Claus is Coming to Town. First, let me start by thanking the office for getting ahead of this issue with these roundtable discussions and with the recent registration guidance on the human authorship requirements. AI is quickly becoming a disruptive technology in the music space and the office's proactive approach is both necessary and appreciated.
There are a few important topics that I hope we can tackle here today. First is the issue of rampant, unlicensed use of copyrighted songs in training data in AI models. Songs and recordings are being used without permission and without compensation to the songwriters and artists who created them. This is not fair use. It is large-scale copyright infringement.
As several panelists have already mentioned, the Supreme Court's decision in the Warhol case clearly supports this view. This must be stopped. Music using AI training must be properly licensed in a free market, which includes the right to op out of licensing if one so chooses. Second is preserving a functional and flourishing ecosystem for songwriters and artists.
The near-term impacts of generative AI are already being felt by music companies which create and license mostly generic background music for brands, TV, and film. And this replacement is a bottom-up process. It'll start with the generic background music, but make no mistake, it will improve rapidly, and it will start to compete with and displace popular performers and songwriters.
Finally, not everything is doom and gloom. There are many positive and mutually beneficial opportunities ahead for human writers and artists, as well as the AI companies, such as human-machine ideation, collaboration, and advanced tooling assisting the songwriting and recording processes. There are also licensing opportunities for copyrighted songs or vocal recordings to be used as training data, if one so chooses. While we embrace the technological change that generative AI brings, we must not lose sight of the humans -- past, present, and future -- who have and who will contribute human authorship to music.
We must ensure that they may continue to earn an honest living from their work, creativity, and talent, without having the fruits of their labor stolen wholesale, chopped up, and regurgitated through an AI model, without consent, compensation, or attribution. The human authorship requirement isn't just a key component of copyright law. It's a key component to actual creativity, innovation, and preserving the long march of the progress of the useful arts for future generations. Thank you.
>> Jason Sloan: Thank you. And Kathleen. >> Kathleen Strouse: Thank you. Thank you to the U.S. Copyright Office for convening these sessions
and the chance to participate today. My name is Kathleen Strauss. I'm the senior vice president of operations for SoundExchange. SoundExchange is the premier music tech organization independently formed in 2003 to build a fairer, simpler, and more efficient music industry through technology, data, and advocacy. Representing the entire recorded music industry, SoundExchange closely monitors, assesses, and advises on the legal, political, and business impacts to the industry of emerging technology. SoundExchange was created for the streaming era, and we have distributed over $10 billion on behalf of more than 650,000 music creators to date.
We continue to embrace cutting-edge solutions, and through real-time data management, we process billions of performances each month and are constantly honing our best-in-class patented matching technology. We process and pay out 90% of royalties within 45 days of receipt. While our matching technology is pivotal to that, working with DSPs, streamlining reporting data, and sourcing data from copyright owners are cornerstones to accurate and timely payments. A globally recognized leader in the music industry, SoundExchange continues to develop and influence worldwide data and technology standards to ensure accuracy and efficiency through a continually evolving digital ecosystem.
Through the use of data, SoundExchange is making it easier for digital service providers to fulfill their obligations to pay creators fully and fairly for their work. We pay out monthly, operating with one of the lowest overall administration rates in the industry. At SoundExchange, our mission is to power the future of music, and for music to have a future, it must be fair to the creators who drive it. As one of the founding members of the Human Artistry Campaign, we believe that creators must be the center of the conversation and that AI tools should be operationalized in a way that protects artists and the value of music.
The music industry is poised to transform once again, and we must anticipate the challenges ahead, both immediate and long-term, so we can maximize opportunities for creators and enact guidelines for responsible use of artificial intelligence. Thank you. >> Jason Sloan: Thank you. Thank you all for those introductions. To begin the discussion, I'm going to hand things over to my colleague, Chris. >> Chris Weston: Thank you, Jason.
So the Copyright Office is interested in learning more about how creators are using and plan to use generative AI in the creation of musical works and sound recordings. So I would like the panelists to expand on this, and in responding specifically, can you please discuss your views on whether there are situations where generative AI is used as a tool as part of a larger creative process that is driven and controlled by a human being? So if you'd use their raise hands feature and I will call on the first person. Alex. >> Alexander Mitchell: Thanks so much, Chris. You know, absolutely. Generative AI isn't just being used now.
I mean, this is not a new thing. It has been used, like I said, for generations. I think the generative the term generative music actually comes from the activities of the artist Brian Nino, you know, Robert from Radiohead, Apex Twin. Many, many examples of artists who have used different types of algorithms in their creative process.
And to take some of these new methodologies and apply them to that creative process, it seems like musicians making music. And I think that's an important filter to apply when people are talking about this stuff. Is this, you know, some sort of AI robot as it is sometimes discussed? Or is this just musicians doing what musicians have always done? I'll give you a specific example from Boomy, where we are really looking at this from the copyrightability perspective on different activities that people might take or might engage in when they use our platform. So, for example, on Boomy, you can generate a song, edit that song, rewrite that song, add your voice to that song, spend hours and hours editing and changing that song.
In our view, that would be clearly quite a bit of human effort that goes into the creation of a song that might be described as generative or might be described as AI-generated. But it's certainly coming from an artist. There's another set of activities that could be called curation, where you've determined a human has used their creativity to determine a set of inputs and then have run that algorithm over and over and over again until the sort of algorithmic system can create something that they want. So those are two types of human labor, two types of work that's going into the creation of the ultimate song, notwithstanding everything that our musicians have done to create these algorithms. And so that's one just one example of an area where you have a new artist or a new musician maybe creating music for the first time and doing something that is fundamentally a musician making music. So I would say that at least the way we look at it and at least the way we design our generative systems, of course, it is part of a larger creative process.
And of course, that's completely controlled by a human being. >> Chris Weston: Okay, thank you, Alex. Kevin. >> Kevin Madigan: Yeah, I'll just sort of piggyback on some things Alex said and also that Shira said in her opening remarks.
And that is that, you know, music creation is an area of the arts where AI technologies have been used for years and are currently used to produce works. And musicians and music producers use them as a tool for a larger creative process. If you think of things like a beat generator or like auto tune vocal tools, these have been used for years. And while they may utilize AI technology, they aren't models that ingest massive amounts of creative, copyrighted, productive works. And that's an important distinction that we should make.
So whereas many AI technologies are used as part of a greater creative process, there are different AI platforms that may be ingesting full songs or sound recordings for the purpose of creating new musical works or sound recordings. And in that scenario, the person sort of prompting the system to generate a song may not be a musician or songwriter. And they may not be really exerting any creative control over the generation of new works. And, you know, then that new work might actually act as a substitute for the songs ingested. So, you know, I guess just sort of follow up on something I said in my opening remarks, I think it's important to recognize when generative AI technologies have sort of a substitutional purpose rather than when a creator uses it as a tool to sort of supplement an already existing work or, you know, the creator manipulates the tool in a sufficient way.
>> Chris Weston: Thanks. Jack is next. >> Jack Kugell: Sorry, I was muted there. I was going to say at this point, most generative AI tools for songwriters appear to be in their infancy, at least in my experience. Songwriters are not quite yet putting in a prompt and receiving a fully baked composition.
But we do know that its widespread use is just around the corner. Anecdotally, we've heard the generative AI is mostly being used for ideation at this stage. But there is concern among songwriters about AI being used, particularly where you don't have a guarantee that the AI you're using doesn't infringe on someone else's work or isn't stealing from human creators.
>> Chris Weston: Thank you. So does anybody else have any responses? Nathaniel. >> Nathaniel Bach: Thanks, Chris. Yes. Just a few other thoughts. You know, I think, again, we're at the infancy of this.
And so this is going to continue to roll out and it's going to continue the conversation. Pro Tools, Ableton, GarageBand. We will get to the point, presumably, at which AI processes will be so folded into the standard suite of products available to songwriters and producers that they may not even know what is necessarily happening when they press that button. It will just become de rigueur.
And so having an understanding along the lines of what Kevin said is to what is feeding into that process on the back end. Where the technology, where the learning, where the magic is coming from and whether that's trained off of copyrighted works without a license is really where the rubber meets the road. Because at some point it's just going to become part of our Microsoft Word, part of our suite of products that are in front of us every day. >> Chris Weston: Okay, thanks.
Rohan. >> Rohan Paul: Yeah, I just wanted to touch upon a couple of tools as an artist. I think there's a lot of AI tools that help with different steps in the process, which used to be quite complicated for people. And I think that helps democratize creation and is a net positive as long as they're sourced ethically.
So these are tools that could help you compose songs or design samples for your songs, write lyrics to your song and even just master your song. I think a lot of these tools open up new ways for people to create music where they would normally otherwise struggle. But I think when we talk about like Music LM, which is like this new model Google released, I can totally see people in the future trying to train a model on a specific artist's catalog.
And I don't see a world where it makes sense for that person to end up having a million songs in this artist's style that they can own and monetize while the artist only has their original 10 songs that they created. So I think it's important for catalogs, more than just mass trained catalogs, catalogs that are specifically trained for a model should be owned by the creator rather than whoever decided to take them and train that model. >> Chris Weston: Okay, thanks.
We've got a few minutes left at this particular topic, if anyone else has any thoughts. If not, we could go to the next topic, and I will ask Jason to ask the next question. >> Jason Sloan: Thanks, Chris.
So we've heard how certain AI models, generally for text and images are trained, and the concerns that creators and copywriters have with models that use their copyrighted works without permission as part of the training process. We heard the similar concerns from several of you all during your introductory remarks. When it comes to musical works and sound recordings, what's your understanding of how current and emerging generative AI models are being built.
For example, are they similarly trained on pre existing musical works and sound recordings, or are there other methods, such as applying musicological rules. Are there any technological distinctions between AI training of musical works and sound recordings as compared to other types of works like text and images? Jason. >> Jason Rys: Yeah, sure. So I think you hit the nail on the head there.
There are a couple different types of training processes that the AI companies are using. Some are using musicological rules. And, you know, as a publishing administrator, zero problem with that. Great. Love it. I support the, you know, democratization of access to music in that fashion. I think where it breaks down is there are a number of companies that are taking in copyrighted music through illegally gotten ways, scraping the web or, you know, whatever data set that some researchers collected.
And they're taking this copyrighted music and they're feeding it into their algorithms, and out the other end, they're producing things that are derivative works based on those copyrighted original songs. So I think there's two buckets to consider and they have different ramifications from a copyright perspective. >> Jason Sloan: Thank you. Ken? >> Kenneth Doroshow: Yeah, I'd just like to add that there's a lot of focus on lyrics, and so really the language models must intake those lyrics, they must take copyrighted lyrics in. So that's not a case where you can ask the question as to whether or not you've just applied musicological rules to it.
You do need to intake those lyrics wholesale. >> Jason Sloan: Thank you. Rohan. >> Rohan Paul: Yeah, I just wanted to touch upon specifically AI singing voices for this and how they're trained. When they're trained on celebrity models, a lot of times they will take existing songs on YouTube and use stem splitters to get just the vocals, and then they'll train a bunch of those vocals so that they can recreate the timbre of that voice.
And in this case, what the technology does at its core is it's taking one single vocal performance and it's converting the timbre into someone else's voice. So when people do that, it's both using that original recording, that reference audio, and it's using -- it's cloning that other person's voice. And I think it's kind of irrelevant where or how much data they use to clone someone's voice. I think if they can clone a singer's voice in an indistinguishable way, that singer should own it, regardless of who owns the catalog of music that it was trained on. >> Jason Sloan: Thank you.
Ken. >> Kenneth Doroshow: So I'm just going to echo some of the points that have already been made here. There are obviously different types of technologies in use here. Distinguishing the sound recording piece from the musical work and lyrics piece just for a second, you know, there's a professed desire by some of these AI companies very openly that they are seeking -- when they're trying to output audio, they want to capture the subtle timbres and dynamics and expressivity of actual audio.
There are some systems that, you know, will ingest and output MIDI files, but then, you know, to get this more rich output, they have to ingest raw audio. And again, you know, not to beat a dead horse, this is a theme, I think, in a lot of the comments here. Those audio inputs have to be licensed.
You can't just simply take them to make this more expressive output. And that's the fundamental problem that we're seeing with all of this. And to echo the points that Rohan made, with the vocal cloning, in particular, we're seeing the proliferation of stem extraction, vocal stem extraction, from copyrighted sound recordings. Those stems are, themselves, copyrighted material.
And you can't just take it without authorization. So, you know, this is a problem we're seeing across the board. There's name, image, and likeness, and right of publicity issues with all of that, too. And hopefully, we'll talk a little bit more about that in detail.
But particularly with respect to the ingestion of inputs for this kind of technology, it just seems, you know, nakedly misappropriative to take this without permission. >> Jason Sloan: Thank you. Anybody else have anything to add? Nat? >> Nathaniel Bach: Yeah. To build on what Ken was saying, you know, I think whenever there's transformational technology and a shift in the landscape, we kind of enter the land grab phase where, you know, obtaining market share is extraordinarily important.
And we've seen this previously with respect to self-driving cars, delivery services, whatever it might be, the Ubers, the taxis, et cetera. And then the profit and sustainable models sort of come second after they've captured customer base. And I think we're still in that phase. And so setting the ground rules while we're in that phase and having an understanding is critical to protect artists. And again, there's nothing preventing companies who are actually scraping, using inputs and ingesting copyrighted works, songs, or recordings from going out and getting licenses. And in fact, we're hearing today from folks like Rohan and others, they're saying, we want artists' consent in order to build our business, in order to make it sustainable and scalable long-term.
And so the question then becomes, well, why couldn't that, and why isn't that the default across the industry as a whole? And it should be. And in fact, Warhol speaks to that, right? The Supreme Court's decision in Warhol speaks to that just a few days ago in which you're talking about the same purpose. You could have two different uses, two different companies, one giving a license, one trying to rely on a fair use defense.
And I think that that opinion points up exactly the concerns that are being articulated today and why a license is necessary in that circumstance. And I think Justice Sotomayor said something to the effect of, you know, why not pay -- why didn't you simply pay Goldsmith a few bucks for a license or whatever it would have cost at the time? That really resonates with me and with MAC here when we're talking about what these companies should be doing as a baseline. >> Jason Sloan: Thank you. Alex, if I may, I want to ask you a specific question based on something you said in your opening remarks.
You had made reference to, I believe you called an original AI framework suggesting, it was somehow maybe different from some of the other things being discussed and I wanted to see if you could elaborate on what you meant by that. >> Alexander Mitchell: It's been a very common question these days, particularly in conversations like this. And I've got a lot of different ways of explaining this and none of them are great.
The best way I would describe it -- and we've heard, you know, some other support for musicological rules-based composition, which, again, I cannot reiterate enough, has a very long history and a very long history of copyrightability. I think that when it comes to our approach, you know, we are taking those -- there are a lot of different ways to solve a statistics problem. As a musician, it kind of pains me to describe the creation of musical work as a statistics problem, but that is a helpful way of understanding some of the different methodologies that can be applied to generating music. You can solve a statistics problem by creating a gigantic black box model. You can also solve a statistics problem by hand.
You can also understand the relationships between notes on a musical level, create algorithms that will, you know, take in those rules and those relationships and generate an output that in our experience very much needs human intervention and benefits from human intervention in order to sound great and sound awesome. Certainly, within the millions of songs getting created on Boomy, the best that we've heard and the things that we are pulling out and supporting as a label and publisher in the most traditional sense, of course, tend to be the songs with the most human intervention and the most sort of, let's call it, you know, vocal quality, the artistry, the human artistry that gets applied. And, you know, to answer the broader question, we pay very close attention, and we've been in this market for several years, to advances in methodologies that, you know, obviously we've seen, right, with LLMs, with things like Stable Diffusion. And, you know, there has been a discussion here of market forces.
Clearly there's also been a discussion of, you know, companies who have jumped into the market, maybe without licensing in other domains, not largely music, but in other domains. Whatever we want to say, there is now a market expectation for prompts, for using musical influences and using natural language to create music. And so for us, we have been very active in doing everything we can to try to obtain licenses, to be able to answer that market demand for prompts. And I think we'll probably spend more time discussing licenses inside of this.
But for now, what I would say is, you know, as one of the leading platforms in this space, we have, you know, wanted to respect copyright. We've always respected copyright. We've answered these statistics problems by hand so far. We would love to work with rightsholders and obtain licenses to answer the market demand for prompts and continue to build what we, again, what we see is a human created work.
I have more thoughts on that, but we'll probably save them for the later questions. >> Jason Sloan: Great. Thank you. So several of you mentioned this in your introductions, but would anyone like to elaborate on the concerns that creators and copyright owners have with various training approaches with respect to using pre-existing copyrighted musical works and sound recordings as part of the training model? Sorry, Jack. >> Jack Kugell: Sorry about that.
Thank you. Yeah, a few things. I mean, songwriters in general, we're extremely concerned about AI developers using their music and works to train AI for a number of reasons. Individual creators have no negotiating power with AI systems developers, some of them being the biggest companies in the world. And because of that, we're kind of at significant risk of being harmed economically by unauthorized use of our works for training and ingestion purposes.
When folks talk about the sound recording being ingested, we've got to remember there's an underlying composition, an underlying work that also needs to be licensed. It's not just teaching someone how to sing like Frank Sinatra because they're licensed to master recording. You can't have a master recording without an underlying composition.
And I think that's something that needs to be remembered and is not always remembered. AI developers using copyrighted music to train -- they're using it to train precisely because it does have value. And the fact that we're using it at all shows that our works should be compensated accordingly.
Unauthorized AI training run the risks of cannibalizing the marketplace, stealing, ingesting copyrighted works and then outputting works based on those. They compete with the work made by humans. And we're going to be at a point in the not-too-distant future where they'll be able to do it cheaper and faster. >> Jason Sloan: There you go. Thanks, Jack. Kathleen.
>> Kathleen Strouse: So to reiterate what I think other people have said today, as a general principle, all copyright creators and owners have the right to determine how their works are used. But for this right to be meaningful, the ability to track how copyright works are used in AI training sets is critical. This data includes identifying not only input works themselves, but their owners, their creative participants, so that proper licensing around that use can occur.
>> Jason Sloan: Thank you. Kevin. >> Kevin Madigan: Yeah, I'll just be brief because I think others have made these points, but I just would say absolutely creators and copyright owners are concerned if an AI system is ingesting and engaging and also copying of sound recordings, particularly if the purpose is to generate a new work that will potentially act as a market substitute for the work that the system trains on.
And if you just sort of think about it from the perspective of a musician or songwriter, their works are potentially being used without their permission to fuel a techn
2023-07-01 16:57
