Copyrightability & Generative AI in Music
Artificial intelligence (“AI”) is everywhere. It promises (or threatens, depending on how you view this technology) to revolutionize nearly all aspects of our lives, and some artists fear that so-called generative AI will make it even more difficult to make a living in the music industry. But before we start ringing any alarm bells, one of the first questions that needs to be answered is: Are songs created using generative AI copyrightable?
Last fall, I sat down with my friend and former music school classmate, Andy Sydow, a Nashville-based singer-songwriter, to discuss this very topic. You can watch the full video here, which is on Andy’s YouTube page.
Can I Register a Copyright for a Song that Was Created Using Generative AI?
Most of today’s publicly available generative AI platforms produce an output using one or more inputs (or “prompts”), often in the form of text. In the music context, a user typically enters a text prompt, and the platform generates an output song. For example, the (purposely silly) prompt Andy used to generate the song in the video was:
This is a song about Adam and Andy. They met in music school at the University of Colorado at Denver. They had several classes together, but their favorite was the Claim Jumpers “trad jazz” ensemble, where Adam played drums and Andy played piano. Since graduating over 10 years ago, Adam and Andy have stayed in touch periodically. And now, they are reuniting for the first time in many years to create the greatest podcast of all time.
Andy entered the above prompt and asked the platform to create a bluegrass song with female vocals.
Is the Output Song Copyrightable?
The Copyright Act protects original works of authorship that are fixed in a tangible medium of expression. In other words, the overarching “elements” of copyrightability are: 1) originality 2) authorship and 3) fixation. For the originality requirement, only a very small amount of originality is required. In fact, originality is often thought of as “little more than a prohibition of actual copying.” Suffice it to say, when it comes to output songs from generative AI platforms, the originality requirement is almost certainly satisfied.
Skipping ahead to the fixation requirement for a moment, this is a similarly low hurdle. All that is required is the work must be sufficiently permanent or stable to allow you to perceive it for more than a fleeting amount of time. The work can be perceived “either directly or with the aid of a machine or device.” So, when a user enters a prompt into a generative AI platform and listens to the output song on his or her device (e.g., tablet, smartphone, computer, etc.), at least two “fixed” copies are made: one on the platform’s servers and one on the user’s device. The user can go back and listen to the output song (i.e., “perceive it”), and the user can even download the song to his or her device, which would be a clear example of sufficient “fixation” for copyright purposes. Even though the song is not embodied by itself in a tangible medium such as a vinyl record or CD, it is sufficiently fixed for copyright purposes.
That brings us to the more interesting question: Are the output songs produced by generative AI platforms “works of authorship”? For copyright purposes, authors must be human. But the question still remains: Who or what is the author when it comes to output songs from generative AI platforms? Is it the person entering the prompt, the platform itself, or a combination of the two? Certainly, if a song is created “autonomously” by the AI platform (with no involvement from any human), that song would not be copyrightable because there is no human author.
But what if a user enters a text prompt? Does that make the user the author of the output song? Based on recent guidance from the U.S. Copyright Office, simply entering a text prompt does not involve enough human creativity to make the user the author of the output work. The Copyright Office has taken the position that there is simply too much of a disconnect between the text prompt and the output work to say that the user entering the prompt is the author of the output work. An example of this disconnect is the fact that for many generative AI platforms, users can enter the exact same text prompt over and over and never get the same output. Contrast that to a musician playing the piano: every time the piano player hits the middle C key, the piano should produce a sound that the piano player and listener both recognize as middle C.
The important distinction here is how the AI is being used. If the AI is being used just to enhance the human expression, it will not impact the copyrightability of the work. But AI platforms that make expressive choices likely will not produce copyrightable works. For example, if a mixing engineer inputs a bass track into an AI platform and asks the platform to EQ the bass to emulate a certain model of amplifier, it is easy to see how the AI is being used just as a tool to modify the bass track without meaningfully altering the human expression.[1] On the other hand, if a user inputs “funky bass line in the style of Bootsy Collins,”[2] the AI platform is responsible for making all the expressive choices to create the actual bass line. The Copyright Office has likened text prompts to instructions given to a commissioned artist. If a customer pays an artist to write a love song about their significant other, it would be difficult to argue that the customer is the author of the song, no matter how detailed the instructions were, because it is the artist’s interpretation of the instructions that actually creates the song.[3] That is the whole point of hiring the artist in the first place, and that is half the fun of generative AI for some users. You never quite know exactly what the output song will sound like.
Now the question becomes, what happens when a human creates part of a work, and the AI platform creates other elements? In these situations, the Copyright Office has typically granted a partial registration for the human-created elements of the work. For example, the Copyright Office recently granted partial registration for a comic book, Zarya of the Dawn, incorporating elements created by a human and other elements created by a generative AI platform using a text prompt. The human author wrote the story and used parts of the story as text prompts to generate images to go along with the text, and the author also arranged the text and images to create the final product. The Copyright Office granted registration for the story and the author’s arrangement of the text and images, but not the images themselves, as they were created by AI using a text prompt. In the musical context, if a songwriter used a generative AI platform to create the lyrics of a song and then composed the music to go along with the lyrics (i.e., the melody, harmony, and rhythm), it seems likely that the Copyright Office would grant partial registration for the copyrightable elements of the music itself (likely the melody and harmony), but not the lyrics.
What if an artist uses generative AI to create a song and just uses that song as inspiration? That should not impact the copyrightability of the song the artist creates because that would be no different than listening to any other record for inspiration.
Although the ultimate impact of generative AI on the music industry is unknown at this point, the Copyright Office has made it clear that their position is unchanged: human creativity and expression is the focus of copyright law. Output songs created by the currently available generative AI platforms using a text prompt are not copyrightable because there is insufficient human authorship. That said, just because an artist uses an AI tool in connection with their songwriting or production process does not automatically mean that the song and/or sound recording are not copyrightable. Artists can receive partial registrations for the elements of a work that they create, and if an AI tool is just being used to enhance the human expression without making any expressive choices itself, that should not impact the copyrightability of the work. The important distinction is how the AI is being used.
[1] There are two copyrights in every recorded song: the underlying musical composition and the sound recording itself. This article focuses on the musical composition, but the result is likely the same when it comes to the sound recording. The AI in our example is not being used to make any creative choices for the sound recording. In fact, mixing engineers would argue that the creative choice as far as mixing is concerned would be the decision to emulate the particular amplifier for its unique sound quality.
[2] Using Bootsy’s name in the prompt raises another issue, namely his right of publicity, which is a topic for another article. Some AI platforms are instituting protections for artists’ rights of publicity by not allowing artist names to be used in a text prompt.
[3] In this example, we are assuming that there is no “work made for hire” agreement in place.
