"The AlgoRhythm Is Gonna Get You"
The Recording Industry Takes Its Shot At LLM Developers
Images, Books and now Audio Recordings
As the saying goes, “I have seen this movie (image/audio recording) before.” AI Image creation tools and those merely predicting text have both become targets of well known lawsuits currently in litigation. Next up - audio generation tools.
Among the most popular AI models producing audio are developed and deployed by the services Suno and Udio. These tools enable users to enter text prompts and generate lyrics, music or both. Here is one of the user prompt created examples in which the user provided lyrics and Udio provided the melody, music and the generated voice. A piece entitled Can’t Stand The Rain.
The Plaintiffs are various copyright owners, specifically, owners of copyrights on audio recordings that claim these two tools trained their models using, in part, their copyrighted works. The remedies sought are
declarations that the two services infringed plaintiffs’ copyrighted sound recordings;
injunctions barring the services from infringing plaintiffs’ copyrighted sound recordings in the future;
damages for the infringements that have already occurred.
The simplicity of the claim is that an AI model that trains on copyrighted works is copying those works in a way the Copyright Act (Act) prohibits. Given the Act was created and modified before the advent of the ability of companies to create such AI models, the viability of such a claim is a grey area.
What About Humans?
My analysis of this type of lawsuit is similar to that regarding the image related lawsuits. How do humans make music, specifically, the kind that the RIAA seeks to protect? Presumably, a musician starts in childhood learning an instrument or taking voice lessons and writing lyrics. They, of course, practice by performing basic scales and other repetitive exercises, but, also, they undoubtedly practice hundreds if not thousands of copyrighted songs in the form of sheet music reading or playing by ear. They eventually perform those songs publicly at small recitals all the way up to Carnegie Hall events. Humans repeatedly play copyrighted music, learn its rhythms, melodies, chord structures, etc. They also repeatedly listen to copyrighted music to learn those same things from that experience.
The problem the RIAA will have here is to obscure this reality from the court. Certainly, the RIAA cannot sue Ed Sheeran by pointing out that he listened to, repeatedly played, studied, learned from copyrighted works. They cannot sue him knowing that he undoubtedly utilized techniques, approaches, chord progressions, etc. learned from listening to, mimicking and playing copyrighted songs as he honed his songwriting and playing ability. Since that lawsuit would be ridiculous, the challenge for the RIAA is to distinguish that process from what these companies did to train their AI models.
The graphic above poses the question visually. What are the AI models doing in their training phase that is not being done or not capable of being done by humans? Since it is obvious humans cannot be sued for creating music after training themselves on, in part, copyrighted recordings, how is it that an AI model is learning differently?
Granted, AI model developers can consume ALL the music, that is, way more content than a human can consume, memorize or learn. That is true. But, different humans have different capacities to learn and memorize copyrighted music as well. That difference in capacity is not grounds for suing some over-performing humans but shielding other less capable humans.
How would you, as the lawyer for the RIAA, make an argument that the AI models are doing something unfairly different than humans have always done when learning how to play music, construct songs and lyrics on the way to composing their own songs?
Is The Same Sounding Thing Infringing?
A little more than a year ago, songwriter, musician and singer Ed Sheeran prevailed in a lawsuit claiming that a four chord progression in one of his hit songs was an improper copy from a song popularized decades earlier by Marvin Gaye among other co-writers. A story about that case is here. Why reference this case? Because, the plaintiffs in that case presented “musicologists” with detailed expert reports. Those reports noted very little difference between the key four chord sequence in Sheeran’s work and that previous work. In response, Sheeran presented himself, and his guitar on the stand and quickly went through multiple past and present hit songs using that same four chord progression. He pointed out that this chord progression predated the creation of the very work that the plaintiffs were alleged Sheeran had infringed. He went further. He showed the jury how other multiple four chord progressions are the backbone of many hit songs past and present. This was a convincing demonstration of the utter re-use of portions of songs that happens now and has happened throughout the decades of recorded music.
The jury sided with Sheeran. Likewise in this RIAA case, in what form would a court find that what these music generating AI model companies did is distinct from what humans have always done to generate new music? That issue includes the AI model creating works that sound very similar to or constructed similarly to existing copyrighted works.
The RIAA’s Best Argument
The RIAA is represented by competent and undoubtedly experience lawyers focused on copyright infringement claims. Their claims in the complaint are basically one of existential threat.
“[T]here is both promise and peril with AI. As more powerful and sophisticated AI tools emerge, the ability for AI to weave itself into the processes of music creation, production, and distribution grows. If developed with the permission and participation of copyright owners, generative AI tools will be able to assist humans in creating and producing new and innovative music. But if developed irresponsibly, without regard for fundamental copyright protections, those same tools threaten enduring and irreparable harm to recording artists, record labels, and the music industry, inevitably reducing the quality of new music available to consumers and diminishing our shared culture.” (Complaint ¶ 3.)
The claims of infringement, however, are conclusory. The complaint restates in multiple paragraphs what the RIAA seeks to infuse into the law as an objective, inarguable given - using copyright works to train AI models is infringement. However, that conclusion is the crux of the issue the courts must decide. The Act does not label the use of copyrighted songs to train AI models as infringement - how could it? The Act never contemplated the creation of such models to begin with. Even later court decisions interpreting the boundaries of the Act have not yet, determined that it clearly prohibits this type of use of copyrighted works. Undoubtedly the Supreme Court will weigh in on this issue as a single decision would likely resolve the image, text and audio versions of this lawsuit.
What About An RIAA Win?
Let’s imagine a world in which every court to consider the issue concludes that AI models using copyrighted works is by default infringing those works. What do AI models do now, specifically those that are currently enabling the production of generated audio? First, they go bankrupt if they are required to pay the minimum cost per infringing work used. Next, other companies rise up in their stead and somehow certify and confirm their models were trained only on public domain audio. Great. However, guess what? As Sheeran’s lawsuit made clear, so many later copyrighted works were re-uses of then public domain audio, sheet music and the like. So, again, these new post litigation AI models will be able to easily produce the same type of audio works that the RIAA is currently arguing demonstrate the models were trained on copyrighted works. In essence (and likely in reality) in the world of music there is nothing new under the sun. The AI models trained on exclusively public domain works will be undoubtedly producing works with chord structures, melodies, etc. that these public domain works contain and, you guessed it, were used in many currently copyrighted works. How would an expert or jury determine that an AI model that produces such seemingly infringing works was actually trained only on public domain works?
Another Way Around The Barn
Music is math. Consider how services like Apple Music and Spotify are able to suggest “similar” songs or put together an “easy listening” or “heavy metal” playlist? Sure, they could use manually categorized recordings from established performers in those various categories. However, the proliferation of music submitted to Spotify and Apple Music on a daily basis make that task impossible to manually maintain. No, these services and others are using math. Counting the duration between notes, the relationship of those notes, the number of notes in a measure, etc. They can reliably predict a song’s genre by using these mathematical. This type of analysis omits the need to use the copyrighted work other than studying these mathematical relationships. Once the model is trained that a typical hip hop song has these following mathematical features, an instrumental Beethoven like song has these mathematical features, etc - What of the infringing claims then?
Conclusion
I agree with the RIAA’s contention that one of the main engines behind creative audio works in the last 50-70 years has been the ability of creators to exploit those works for financial gain. However, it is also true that the Venn diagram of what those copyright holders did to develop the capacity to generate that music and what AI models do to develop the capacity to generate that music are nearly completely overlapping.
I remain sympathetic to the one visual artist mentioned in a previous post. She had worked years to develop a particular style of design related to animated characters. Then, in less than a few days, some AI developer created a model and fine-tuned it on her works and that model created hundreds of new works in that artist’s style. The theft of all her time dedicated to the development of that style seems like it cries out for an injunction and damages.
But consider someone in the late 1800s who had spent his entire life developing the skill to manually produce perfect horseshoes. Then, some young entrepreneur worked with a metal smith to create a machine that could stamp out that same perfectly designed horseshoe in huge quantities pushing the price per horseshoe downward so far that the manual horseshoe maker has no reason to continue his work. Would it have been good for a court to step in at that point and insist that the maker of that horseshoe machine stop using the machine or pay the manual horseshoe maker some fee per horseshoe, etc? While not a perfect analogy, AI models are doing things humans have always done, making images, authoring text and creating audio recordings (and video of course). They are creating those items relying on what they have learned from other copyrighted works. They are creating those items in the same way that humans have always created such works — learning, mimicking and modeling their creations after all those that came before them. It is definitely a zero-sum problem for the courts and one that inevitably will land with the U.S. Supreme Court. But, unless the RIAA can persuade a court that what AI models are doing is somehow divorced from what humans have always done to develop the ability to generate new music, the ultimate decision may be that AI is more like humans that we all realized - and very quickly.