Introduction
Many forces go into creating music generated with artificial intelligence (AI music) – some deceivingly authorial, some surprisingly not. I will analyze the influences of AI music, as well as distinguish the authorial contributors who must be recognized as the primary forces of creation. This work is essential to understanding if and when these compositions can be owned. I argue that AI music: 1) has many influences, but only 3 artistic contributors – the model, the data, and the feature producer; and 2) should mostly be in the public domain, but in some special cases can be owned and profited from. These conclusions are not restrictive to current copyright laws, since little exists for the problem case; instead, they should be used to inform ethical considerations that should be undertaken before writing adjustment laws.
AI music is generated sequences of sound composed by an algorithm that artificially learns patterns, melodies, genres, or other tonal factors. Typically, the model used to generate these compositions are generative adversarial networks (GANs). GANs function as a minmax two-player game where one generative model G captures a data distribution, and a discriminative model D estimates the probability that a sample came from the dataset rather than G. Additionally, D and G are defined by multilayer perceptrons and the entire system can be trained with backpropogation (Goodfellow et al., 2014). By feeding this system with a dataset of melodies and discriminating its probabilities of being captured, the GAN can recreate styles of the given dataset.
Figure 1: Diagram of how a GAN’s generator (G) and discriminator (D) function (Tensorflow, 2022).
Generative music as a broad technique has been pronounced in art forms before. Brian Eno, a prominent artist and promoter of the style, used the term to describe any music that is any set of sounds created by a system that is ever-different and changing (Eno, 1996). Art forms under this style have led to music generated by GANs specifically. Other iterations of this style can be seen in algorithmic compositions and live-coding. Algorithmic compositions are defined as any music that has been created with the use of an algorithm. The algorithm can be used as an assistant for physical creations, such as cycling through a certain set of scales on a trumpet, or it can be used for digital creations, such as manipulating oscillators to create rhythmic frequencies (McLean & Dean, 2018). Typically, algorithmic composition is conducted without human intervention; however, live coding offers a human-in-the-loop approach. Live coding is a method of music production that combines just in time interactive programming with human improvisation skills (Wang & Cook, 2004). It is a technique for performing artists to work in the mix by producing visuals and sounds controlled by their programming environment. The described techniques are foundational precursors of AI music because they have set the landscape for problems when computers are introduced to the music composition process.
The use of prior datasets in music creation is similar to a popular technique in music production – sampling. As a broad term, sampling is the process of taking samples of something for analysis. In terms of music, it is a technique used to digitally encode music or sound and reusing a part of composition or recording. These definitions both give insight into how sampling is used and its relationship to parts-as-a-whole. Parts-as-a-whole will be used in this essay to describe the necessity of a part for the entire composition of a piece. An example of a part-as-a-whole is Jay-Z’s use of “A Hard Knock Life” from the musical Annie in the composition of his track, “Hard Knock Life”. The musical’s song is central to Jay-Z’s song; it is a prominent feature and therefore a whole component embedded within the rest of the song – a part-as-a-whole. When music samples others work and uses it as a part-as-a-whole, the song becomes liable for infringement.
Sampling is a contested art form due to copyright laws. Legally, sampling is defined as the act of integrating sounds embodied in a pre-existing sound recording into a new musical work (Cohn, accessed 2022). This means that any pre-existing work such as a melody, chord composition, or any sequence of sounds can be subject to ownership and copyright. Infringement on these pieces results in expensive legal battles, million-dollar payouts, and a loss of artistic integrity. Some example cases show how this legal condition is just. In the case of “Amen Break”, a heavily sampled loop in popular music from the 1969 track “Amen, Brother” performed by Gregory Coleman, the lack of royalties left the original creator to die homeless and broke despite the immense popularity of his composition (Brown, 2020). However, other battles highlight a vagueness in the law.
The number of melodies in the Western scale is a series of compositions using only 12 notes. Lawyer and computer scientist Damien Riehl has taken upon a task to generate every single possible combination of notes to create melodies and wants to release them in the public domain such that no one can claim copyright ownership on them (Riehl, 2020). This is a task that demands generative models exhausting all known melody constructions and putting them in a physical format, a hard drive, so that they are automatically copyrighted. The claim here is that since all of these melodies are created and put on a physical medium, Damien should technically own them. It is a point of contradiction – if this is the case, then no one would ever be able to make music using the Western scale again. This is a hyperbolic claim to show the weaknesses of copyright; the true point of the scenario is to put all of the melodies in the public domain.
Is copyright law in music just or not? To answer this, we must look at an important distinction between both cases – the respective levels of abstraction. In the case of “Amen Break”, artists would loop the clip or insert it directly into their songs such that it was an obvious motif and the use of it in a track would be immediately recognized. In the case of melodies, a series of notes can occur in any fashion and be altered such that they can become new every time it is composed in a different setting. This part-as-a-whole/generative dichotomy becomes a distinct problem when analyzing AI music, which copyright lawyers have been unable to distinguish. The question is not of justice for original creators, it is about the level of abstraction one can take before something can be owned and who is its distinguished owner.
Influences
Who owns AI music? It is a question that has frazzled copyright lawyers and artists alike. Some think the obvious answer is the AI, but this is a poor anthropomorphization of technology and impossible due to copyright under the United States (US) law claiming only humans can have these attributes. As such, US law defaults such that AI artworks all fall under the public domain (Bruch, 2021). Following, the US Copyright Office has claimed that AI-created art cannot be copyrighted (Zhang, 2022). However, in the United Kingdom (UK) legislation says that AI art falls to the individual or group of individuals who created the software or made the arrangements necessary for the work to be produced. The amount of work put into the art will determine who has rights over it (Bruch, 2021). These differences lay a foundation for the intricacies of influences that lay within AI music.
A quantization of influences must be made to analyze the division of labor in AI music. To do so, I will walk through the general process of creating a piece of AI music to highlight various degrees of influence within a final creation of a piece of AI music. First, a dataset must be created. Its composition determines the initial level of influence. If the dataset is of one sole creator, then that creator has an obvious influence – it is parts-as-a-whole sampling. If it is of many creators, such as to distinguish a genre, then the creators have less of an influence and can be regarded as a style. The person that puts together the dataset is another obvious influence, but they can be regarded as a kind of playlist creator or technical disc jokey. This person’s influence is a contribution, but it is not a primary one because they have little impact on the full output. They are a teacher of sorts that the later algorithm learns from; essentially, they are like my primary school teacher who taught me how to write – this person should not be credited for this essay I am currently typing, but they did help me along the way to my current understanding of linguistic flow.
Next, we have the model. The programmers of the GAN is definitely a contributor to the whole piece. What about the programmers of the libraries and platforms that helped make the algorithm? This again goes back to the primary school teacher analogy – these people absolutely contributed to the creation of a GAN model that can create AI music, but they are not primary contributors due to the degrees of separation from the final product. The ownership of the GAN gets a bit more complicated when you start looking into open-source programs versus proprietary software. Organizations typically create these products. Very rarely a single person goes through the process without any help. Regardless, the ownership of the model is contested.
Finally, we have the feature producer. This is the person that manipulates all the features and inputs of a model, such as the weights of a GAN or the styles of genres they want to see mimicked. They also contribute to when the project is done such as the number of iterations a piece should go through until there is something clearly usable or recognizable as a track. This person is in fact a primary contributor because their influence is essential to making the creation.
In conclusion, there are many influencers over the creation of AI music. The first important contributor depends on the type of problem the GAN wants to solve. If it is to create a fake Three Six Mafia album, for example, then Three Six Mafia is an important contributor. If it is to create music in the style of Memphis rap that uses Three Six Mafia along with other artists such as Young Dolph and Moneybagg Yo, then the influences are less detrimental to the system as it would be picking up on a style instead of a single direct artist. The engineers of the GAN are also major contributors to the work, as the output would never be created if it wasn’t for the developers of the system. Lastly, the feature producer is an important contributor to the final output because they are adjusting and manipulating the usage of the system to create its final output. This person is adding insurmountable creative input to the overarching system. These two or three influences should be noted as primary creative forces within the artistic process of creating AI music. As such, they are authorial influences of the outputs of AI music.
Free & Fair AI Music
According to US Law, AI music defaults in the public domain and cannot be claimed by copyright. This seems a bit odd because the decision resembles fair use claims about the nature of this type of work. In this section of the essay, I will go through all four factors of Fair Use as it relates to AI music. I will also discuss the nature of open-source software, its relationship to free software, and why these ideas are important in the future development of creative AI technologies.
Fair Use is the doctrine that excerpts of copyright material may be quoted verbatim for purposes of criticism, new reporting, teaching, and research without the need for permission or payment to the copyright holder. From the perspective of AI music, this implies that copyrighted material may be used under the guise of research purposes. However, these creations must follow the four factors in the Fair Use Balancing Test (Crews, unknown year). The first factor discusses the purpose and character of the use, which implies that a copyrighted use can be used for the previously stated purposes specifically for our case – research. It also states that courts favor “transformative” uses such that they are not mere reproductions. For AI art, this is an easy pass due to the nature of transformation embedded within the model. It is the entire purpose of the creation; to be a transformative new work based on copyrighted work. The second factor is the nature of the copyrighted work, which essentially states that the use of a work must already be commercially available such that it is originally published by the original author. Since AI music only uses songs, melodies, tones, etc. that have previously been created, this is no problem. The third factor is the amount or sustainability of the portion used, which implies that the more copyrighted material you use, the less likely you are within the bounds of fair use policy. Since AI music typically samples work, as in using snippets of songs or musical characteristics, it is permissible under this rule. However, when the sample used is a prominent feature, such as our parts-as-a-whole analogy, it may be less likely to fall under fair use. For example, a song based on the entire discography of an artist. The final factor is less straightforward than the others and introduces more prospective, unknown tendencies. It is regarding the effect of the use on the potential market for or value of the work. In some cases, AI music elevated the market value of an artist, but in others, it diminishes the market value; that being said, the future is unknown and we must follow reasonable trains of thought to understand how these situations may play out. It also depends on the view of the original creator. There is not an easy answer here for AI music and research contexts in general. Potential market value change is difficult to prove in these cases and is thus left as an unknown. Overall, AI music fits neatly into the four factors of Fair Use policy. The decision for AI music to be under the public domain is just in this case, but it is conditional on no profit being made from any part of the process.
The relevance of open-source software is central to the role of AI music development. Many tools for this generative technique are easily accessible online, such as Open AI’s JukeBox and Google Magenta’s GANSynth. Open-source technologies allow for more contributors to add development opportunities and raise more flags of awareness for bugs or complications. By making these technologies easily accessible and usable, artists can produce more products in increasingly novel ways.
Figure 2: Distinction between free software, open-source software, freeware, and public domain software (Moqod, 2019)
However, open-source may not be enough. Richard Stallman, a prominent figure in the free software community and writer of the GNU Manifesto, describes the four essential freedoms users should have as the rights: “(0) to run the program, (1) to study and change the program in source code form, (2) to redistribute exact copies, and (3) to distribute modified versions” (Stallman, 2013). This describes different sets of standards or freedoms than open-source software. They overlap, but they are not the same. Free software promotes sharing and cooperation; it is an ethical imperative to respect user freedoms and promote the described set of principles. Open-source software, however, simply focuses on how to make the software better. Both are steps in the right direction compared to proprietary software, which aims to keep software closed and immutable.
The difference between open-source and free software is an important distinction to be made due to the mutability and sociability of the code. Concerning AI music, this means that the software of the model should be accessible to the public because it follows the restrictions required by fair use. These philosophies need to be embedded within the culture surrounding AI music because it will result in a more ethical foundation.
It should also be stated that ideally, the artists used to make AI music should be contacted or asked for permission. Because artists can be extremely difficult to reach in this respect, it might not be possible. That being said, they need to be contacted out of respect for the original creators.
Owning AI Music
Now that we have analyzed what it means for AI music to be free and fair, I will contrast this with a counterargument of when AI music should be owned. Because the three authorial contributors to AI music are so important to the creation of a new medium, there should be ways for this artistic labor to be profited from because artists are endowed with the fruits of their labor just as anyone else. First, we will analyze the question of what it means to own something and which authorial contributors can own parts or wholes of their product. In contrast to US Law, I argue that artists have creative rights in this new world of creation.
Once again, our authorial contributors are the feature producer, the programmer, and the original artists that the AI art is based. However, this outlines a problem: only one person should be able to own the final output of AI art. It is easy to say that those people can own parts of the AI, such that the programmer owns the model, the original artists own their music, and the feature programmers own their customization. None of this tells us who owns the final output, though. To figure this out we must dig into the idea of ownership and analyze at each step who must request what of whom.
Ownership is the ability to possess something and therefore the ability to sell something. The person that owns the output of the AI music, the final song and creation of the process, should have the ability to own it. I argue that, if the model is open source or free, then the last person to put their hands on the product – the feature producer – should own the final output. That being said, the originator of the chain of creation – the creator of the content used to style the output – must be contacted and offered compensation.
Ethical considerations to be undertaken for this to function must be the chain of command and conversations that all parties interact with. In the case of the Amen break, an example of the parts-as-a-whole idea, the artist was sampled to an exorbitant degree but never received any pay. So, because of this, it is imperative that the original creators of the content being made are contacted and offered compensation for their work.
To own AI art, the following considerations should be taken:
- If one person from start to finish created the entire creative system, (i.e., the same person plays all three authorial roles) then they alone own the rights to the output.
- Consent and compensation must be given to obvious influences, such as Three Six Mafia in a generated Three Six Mafia song
- If the model is creating something in the broad style of a genre or theme, then the creators of the original content in the dataset must give consent.
It is important to note that these considerations must be applied to all models of creative art, not just music, but is specifically discussed as music in this essay because it is a simplified case analysis. The creative laborers behind the process of AI music should be able to make a profit because their effort is worth money. However, if it is to be shared in the public sphere for monetary gain, then other considerations should be taken such as outlined in Part 2.
Conclusion
In conclusion, this essay has outlined several necessary ethical steps to take when an artist wants to own the output of their AI music. It analyzes the contributions each party makes and explains who is an authorial contributor and who is not. Additionally, it discusses the cases of when a piece of AI music should be owned and when it should not, the ethical considerations to undertake in each case, and some history behind this style of music and the ideas of open-source and free software. The entire process of creating AI music is complicated and nuanced. It takes many different moving parts to create an AI-generated output, and the individuals contributing to this art form must be considered and compensated.
References
Cohn Legal, & Cohn, A. (n.d.). Music sampling and copyright law: What’s legal? Cohn Legal, PLLC. Retrieved May 6, 2022, from https://www.cohnlg.com/music-sampling-copyright-law-whats-legal/
Brown, H. (2022, April 13). The Amen Break — how an obscure 1960s B-side became the most sampled song in history. Financial Times. Retrieved May 6, 2022, from https://ig.ft.com/life-of-a-song/amen-break.html
Bruch, N., 3, P. J., & Nicholas Bruch Legal Analyst Nick is a senior analyst for the legal industry. He has worked with law firms of all sizes in developing and developed markets on issues related to strategy. (2021, July 4). Who owns the right to ai created work? The Courtroom. Retrieved May 6, 2022, from https://thecourtroom.org/who-owns-the-right-to-ai-created-work/
Crews, K. D. (n.d.). Cul – Main Content. Fair Use | Columbia University Libraries. Retrieved May 6, 2022, from https://copyright.columbia.edu/basics/fair-use.html
Deep convolutional generative Adversarial Network : Tensorflow Core. TensorFlow. (n.d.). Retrieved May 6, 2022, from https://www.tensorflow.org/tutorials/generative/dcgan
‘Generative Music 1’ pieces are by Brian Eno; Title was published and distributed by SSEYO; Webpage content Intermorphic Ltd. (n.d.). Brian Eno’s “Generative music 1”. Wotja 22 by Intermorphic. Retrieved May 6, 2022, from https://intermorphic.com/archive/sseyo/koan/generativemusic1/history.html
Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2014, June 10). Generative Adversarial Networks. arXiv.org. Retrieved May 6, 2022, from https://arxiv.org/abs/1406.2661
Moqod. (2021, October 19). Understanding open-source and free software licensing. Medium. Retrieved May 6, 2022, from https://moqod-software.medium.com/understanding-open-source-and-free-software-licensing-c0fa600106c9
The Oxford Handbook of Algorithmic Music. Alex McLean; Roger T. Dean – Oxford University Press. (2022, May 3). Retrieved May 6, 2022, from https://global.oup.com/academic/product/the-oxford-handbook-of-algorithmic-music-9780190226992?cc=us&lang=en&
Riehl, D. (n.d.). Damien Riehl: Why all melodies should be free for musicians to use. Damien Riehl: Why all melodies should be free for musicians to use | TED Talk. Retrieved May 6, 2022, from https://www.ted.com/talks/damien_riehl_why_all_melodies_should_be_free_for_musicians_to_use?language=en
Stallman, R. (2013, September 28). Why free software is more important now than ever before. Wired. Retrieved May 6, 2022, from https://www.wired.com/2013/09/why-free-software-is-more-important-now-than-ever-before/
Wang, G., & Cook, P. R. (2017). 2004: On-the-fly programming: Using code as an expressive musical instrument. A NIME Reader, 193–210. https://doi.org/10.1007/978-3-319-47214-0_13
Zhang, M. (2022, February 23). AI-created art cannot be copyrighted, US Copyright Office says. PetaPixel. Retrieved May 6, 2022, from https://petapixel.com/2022/02/23/ai-created-art-cannot-be-copyrighted-us-copyright-office-rules/
vemiotics 