Acknowledgments
Firstly, I would like to thank my supervisor, Andreas Waaler Røshol for continuous guidance and inspiration throughout my journey of researching and writing this thesis. I am also deeply thankful to Rachel Meddings for inspiring the topic of this thesis, and most importantly, for your friendship.
Thank you to the artists for letting me take part in your creative process, as well as for sharing your thoughts and opinions with me. Some great songs came out of this because of you:
Raine St. Frank
Torfinn McKenzie
Ida Sofie Laland
Thank you to my dear family and friends for all the support and uplifting words during the process of writing. I could not have done it without you. Lastly, thank you to Halvor, for your continuous support, and patience and for making me coffee when I needed it the most.
Abbreviation
AI – Artificial Intelligence
Back catalogue – Lyric portfolio
ChatGPT– An Artificial Intelligence website made by OpenAI
CLG - Lyric Generator (AI)
DAW – Digital Audio Workstation, software used to create and record music
Demo – A demonstration of a song
Logic – Logic Pro X, a software program to create and record music.
Scaler2 - Musical tool. In this thesis used to generate chord progressions(AI)
Session or writing session – In this thesis, a timeslot where we create music
Teams – Microsoft Teams is used as a video call for collaborations and interviews
Abstract
In this thesis, I explore and research songwriters' experience of using artificial intelligence as a songwriting tool. To do so, three participants have explored the use of two different artificial intelligence tools in their songwriting session, in collaboration with me. One of the tools is used to inspire lyrical writing and one is to create a starting point for chord progressions. To gather data, I have had three collaborative songwriting sessions, where the artists and I use the tools whilst writing a song for their artist project. As well as the sessions, there was held six qualitative interviews, two per artist. The first one was for documenting their prior experience and expectations. The second interview was held after they had received a produced demo. This was used to ask about their experiences writing songs with artificial intelligence. We also talked about the song related to ownership. The thesis questions whether artificial intelligence will make one more productive, too productive, or even if the tools can be perceived as distracting.
In addition to using Artificial Intelligence in three different songwriting sessions, I have researched the ethical dilemmas with using AI, such as copyright infringements, and discussing options for opting out of having one’s work be a part of the training of an AI.
My personal journey highly inspires the thesis through using artificial intelligence for my musical project. By using the same process as the three participants are taking part in, my debut EP, Sugarcoated, was created.
This thesis is approved by SIKT (816719).
1 Introduction
In the spring of 2022, Rachel Meddings and I had a writing session, where we would write music for my project. At that point, Meddings had landed a job with Music Tribe, and related to her work, we used her AI for lyrics and a second AI for finding chord progressions. We enjoyed this process so much, that we ended up having multiple sessions which led us to release a five-song EP in February 2023. The EP solely consists of songs written in this format and was named Sugarcoated.
This was an inspiring process and made me want to continue the research of using AI in songwriting with other artists. The feeling of being inspired is related to what Geir Kaufmann writes, (2015, p. 117) “In the group, the diversity of abilities, knowledge and skills will mean that you can positively complement each other”. This both goes for working with Rachel, but also for the contribution of the artificial intelligence tools added to the process. The inspiration present in the writing process of writing the EP also relates to a creative flow and an intense focus, which will be further explained later in the thesis.
Questions relevant to my research focus on the artists’ prior experience working with AI, their thoughts around artificial intelligence and songwriting, and whether such songs will lose any human aspect when used. Will the AI trigger inspiration or will it be distracting? I am also curious about how the artists compare their AI songs to their back catalogs.
The main research question which is discussed in this thesis is:
How do songwriters experience the use of AI in their songwriting process?
1.1 The research question
How do songwriters experience the use of AI in their songwriting process?
Firstly, when discussing the word songwriter, this thesis will be focused on the contemporary view of what a songwriter is. This includes lyrics, meaning the words of the song, and the melody. To create lyrics and melodies is often referred to as toplining (Bennett, 2011). A contemporary view also includes the arrangement, form, and often the production of the songs. The production is meant to “provide the musical value of the song, all while remaining in the background” (Frith & Goodwin, 1990, p.192). A further explanation of what it means to produce music can be read at the end of this chapter. Simply put, it means to create or further develop arrangements, and add sonic features and effects. A songwriter can therefore include both being a music producer and a top-liner. A process that finally results in a finished song, and in this case, a pop song. The lyrics of pop songs are stories set to music, and are inspired by various genres, as explained by Frith and Goodwin (1990, p.188) below:
From poetry it [pop songs] borrows the importance and autonomy of certain key words, as well as the use of meter, verse and repetition; from lyric theater it borrows the singer’s direct appeal to the audience to share feelings expressed in the first person; but perhaps it owes most to the novelette in the way that it almost invariably tells a story, set out in a few words, concerning the relationship between two or three individuals
How do songwriters experience the use of AI in their songwriting process?
An experience is described by the Cambridge Dictionary (2024) as “knowledge or skill that you get from doing, seeing, or feeling things, or the process of getting this”. This thesis will look at how songwriters feel using artificial intelligence while writing songs. The experience of writing songs can differ from time to time. The songwriters that are participating in this thesis, have yet to try writing songs with AI tools. This will therefore be a new experience and result in knowledge and skills of how to use AI in songwriting.
How do songwriters experience the use of AI in their songwriting process?
“To put something such as a tool, skill, or building to a particular purpose” is the Cambridge Dictionary’s description of the word “use”. For this thesis, one is putting the AI tools to the particular use of writing songs. The focus will be on the use of a lyric generator and a generator used to make chord structures. How do the songwriters react to the use of these in the creative process of writing a song? Will the use of it be distracting, interesting, inspiring? This will be further explored throughout the thesis.
How do songwriters experience the use of AI in their songwriting process?
AI, short for artificial intelligence, is a type of “computer technology that allows something to be done in a way that is similar to the way a human would do it” (The Cambridge Dictionary, 2024), which is an important part of this thesis. By reading the description, one can think, if AI thinks like us, why not just do the job ourselves? Even though the machine is simulating something a human can do, it shortens the process in many cases. AI can produce a lyric in seconds, whereas a human undoubtedly will need more time. However, the quality of the writing and how it feels to sing the generated text is a different question. It is important to apply that the AI utilized for this thesis will strictly be used for inspiring lyrics and chord progressions. The lyrics generator used in this research project is not able to and will not be used to write a full lyric. The two AI’s are solely used to trigger inspiration for further writing, not to replace songwriters.
How do songwriters experience the use of AI in their songwriting process?
My research question attains the word “their” because we firstly are writing music for their projects. Secondly, it is their experience of using AI that is the main focus of the thesis. The artists participate in one writing session each where they gain the experience of using artificial intelligence concerning lyrics and chord progression. Their expectations and experience will be documented with one interview before and one after the session. They will be asked about their process and how AI did or did not affect the musical process and result. Using AI in their creative process will be new to the participants, but also I will be a new factor. As I am in the session with all three artists, the musical work is considered a collaboration, which will affect the process.
Not only will I have an impact on the songs that are written, but I have also taken the job upon myself to produce demos of the songs. I feel it important to describe what is referred to by using the word produce throughout this dissertation. Through different times and genres, the word holds different meanings. The thesis will have a contemporary viewpoint of the word producing. Frith and Goodwind (1990, p.3) explain it as such:
At one extreme the producer’s job is to keep the artists happy; at the other it merges into sound engineering. […] a producer is «perhaps as an expert listener able to focus on the larger context as well as the details of the recording as it develops.
A recording holds an important role in the genre of pop, which is the umbrella genre of what this thesis will be operating. Similar to what was mentioned about producing, also the recording holds different sentiments in different genres. “[…] classical performers tend to see recordings as in essence reproducing concert performances, whereas for pop performers the record is, so to speak, the primary text” (Frith & Goodwin, p.2). All the songs made concerning this thesis find themselves in the pop landscape. The recordings, even though only demos at this point, are the primary representation of the songs.
A demo is short for demonstration and can be viewed as a semi-developed production of a song. “The ideal song demo is one that clearly conveys your song’s hit potential”, Jason Blume (2022) writes. To be able to do so, it was important for me to find production references to the participants. This indicated referencing one of their already released songs and having information on who they are inspired by. This made it easier to find a direction for the production of the recording.
1.2 Background for the thesis
From 2016-2020 I lived in Liverpool, England, where I took my bachelor’s degree in Songwriting and Performance at Liverpool Institute for Performing Arts. During these years, I got to know songwriter and now research music creation specialist, Rachel Meddings, who later became one of my co-writing partners. After graduating, Meddings took her master’s degree in Songwriting, where she wrote a thesis called The AI Songwriter: Coding a Co-Writer (Meddings, 2022) about the audience perception of music made with different AIs. Amongst several AIs, she also created and used her partially homemade AI which is developed based on ChatGPT 2’s technology. The AI is limited to her back catalogue, meaning a collection of all her written lyrics. In that way, she has made an AI that writes lyrics similar to how she does. She called the generator “a machine learning lyric generator trained on 18,000 of my own lyrics”.
In the spring of 2022, we had a writing session, where we wrote music for my project, Celine Lyng. At that point, Meddings had landed a job with Music Tribe and wanted us to use her lyric generator as part of an assignment for work. In addition to the lyric generator, we used a second AI tool for generating chord progressions. Initially, I was sceptical of the thought of having a robot write lyrics for me. I viewed it as if I was being replaced and the question if a master’s degree in songwriting was about to be for nothing. However, I quickly learnt that this lyric generator was not able to replace a songwriter, but to inspire one further. We ended up having multiple sessions like this, using an AI for lyrics and one for chords.
This experience was inspiring for us both and one song turned into five. This led us to release a five-song EP named Sugarcoated, which was released in February 2024. This was such an eye-opening experience, and I was motivated to continue this work with other artists. With permission and guidance from Rachel Meddings, we created a personal lyric generator for me to use for this master project. This lyric generator was named CLG, short for Celine’s Lyric Generator and is only fed my personal back catalogue. By using this lyric generator and a chord generator, this thesis will consist of researching other artists' experiences of writing with AI.
1.3 Limitation
To evaluate the use of AI in a songwriting process, I will first address the limitations that were needed to execute the assignment. Firstly, there was a conscious choice to only focus on two artificial intelligence products. I saw this as an effective way of using AI for both the lyrics and the chord patterns of the song. It can make the session less overwhelming for the artists. I worry it would feel overwhelming and more distracting than helpful if met with too many new tools. Using two tools that have different focuses makes it easier for the participants to separate them and evaluate how they feel using them. In the process of writing a song, the chords can be seen as a starting point and a base for the arrangement. Therefore, I wanted to use an AI to inspire this process. This AI makes chord structures effortlessly and with the option to alter the outcome. The second AI that is used is an AI which is meant to inspire lyrical writing.
Another limitation that was put in order was the number of participants contributing to the thesis. Initially, the plan was to have 4 to 5 people participate to get a broader view of artists' reactions to using artificial intelligence in their writing of music. However, I decided to only have three artists participate in the study. The reason is the limitation of time. This thesis holds two interviews per artist, one songwriting session per artist, and lastly, the additional work of producing and mixing these songs. It had to be decided to limit the number of participants, as the different processes are highly time-consuming.
A potential weakness in doing the research for the thesis is the requirement of creativity. To have a writing session with other people can be challenging, and one’s creativity can be affected by various factors. Henriksen et al (et al., 2020) write “In a world of awashed distraction, stress, and often distress – […] can affect creativity and well-being”. It is challenging to be distracted and stressed over something whilst trying to create a new song in collaboration with someone. Sleep is also important in terms of focusing on a creative task like songwriting. Killgore (2010) states that “creative, divergent, and innovative aspects of cognition do appear to be degraded by lack of sleep”. Another point that feels important to bring up is comfortability and chemistry. To be able to create, in this case, music together, a good chemistry between the collaborators can improve the experience. Good chemistry can also make it easier to open up to the co-writer. Cole (2018) writes:
Be prepared to get comfortable with other writers as you’ll want to share the details of your experiences. Holding back information that a writer may need to paint a picture holds back the impact of your song.
This is relevant to both the participants and me. One can argue that if the sessions are affected by one or several of these factors, it can be hard for the participants to understand or acknowledge why this happened. Instead of seeing the session not going well because of creative hindrances, it can instead be blamed on the AI.
Both me and the artists will be contributing to the songwriting. However, during the sessions held on Teams, it will be my job to control the software. Their ideas will be expressed vocally and executed by me. The roles we take for this thesis are conscious, but not limited to the idea. The choice of me being in control of the software has several reasons to it. Firstly, I am used to using Logic Pro X and Scaler2. Second, the CLG is not released publicly which restricts the ability for other people to access it. Lastly, as it is my research, it felt natural to be the person who has the main control of the session. This can be a limitation.
Firstly, not all artists have a large musical production vocabulary. This means it can be hard for artists to vocally tell a producer their production ideas. Another point is that the one who is not in charge of production can feel passive in the session or even bored while waiting for sonic features to be added. With this in mind, the sessions for this research project were used to make the lyrics and chord pattern of a song. The production was mostly talked about, and the artists gave me references that were similar in style to the song. Therefore, the participants were active during the whole session, and I produced the demo after the session, without them.
The session with Torfinn was different in the way that it was held in person, which allowed us to both contribute to using the software more equally. He was also more involved in the production. His guitar playing is his musical signature, and it was only natural that he recorded those parts himself.
This thesis researches how songwriters experience using AI in their songwriting process. The artists will therefore participate in writing sessions, to experience creating songs with AI. To collect data, there will be held one interview with the artists before the writingsession, and one after they received a demo of their AI song. To write this thesis, I have had to make decisions and acknowledgements regarding limitations. These limitations are the number of participants and AIs, as well as the limitation of time and how easily creativity is affected.
2 Theory
I have chosen to link the thesis to the field of popular music research and will use relevant theories from this field. In this chapter, I will present what I mean with inspiration, as well as present literature on topics that are considered relevant to this thesis. Artificial Intelligence is at a stage where it is developing incredibly fast. It is therefore important to note that, there is a lack of relevant and up-to-date books on the subject. Before the books are published they will, in a way be outdated. This is something Per Kristian Bjørkeng (2018) had to acknowledge whilst writing the book “Artificial Intelligence: The Invisible Revolution”:
With such an insane pace of development as we are experiencing around us today, every single day I receive a handful of new results and sensational projects that should have a place in the book.
This thesis will therefore have theory regarding AI mostly based on articles and newer books such as Artificial Intelligence and Music Ecosystem (2023). When researching AI’s history and its prior development and songwriting, books will be relevant.
2.1 Inspiration
Inspiration and feeling inspired are words often used in creative processes and have already been and will be used going forward in this thesis. As inspiration can imply diverse assumptions, I will therefore explain what is meant by using these words in this research process. Personally, inspiration is a positive feeling of creative flow and an intense focus. With creative flow, one feels as if one becomes one with the writing process. A creative flow, to me, feels like one has been building up a huge amount of impressions and triggers, and suddenly one puts it all into a creative process, where in this case it is creating music. The word was first used by Csikszentmihaly. Cook sums up Csikszentmihalyi's opinions regarding flow (Csikszentmihalyi, 1997, as cited in Cook, 2018, p.6):
Csikszentmihalyi coined the term “flow” to refer to the intensively positive, almost dream-like experience that results from the combination of a challenging real-time task and the ability to meet that challenge, and the term has become part of musicians’ vernacular. Like sex, flow is a for of intrinsic motivation, making creative practice rewarding in and of itself.
By an intense focus, I, personally see it as diving completely into the work, which in this case is the writing of music, that one forgets anything else. It is a positive experience that feels liberating (Miranda, et al. 2012):
Being able to channel all that random energy that’s flying around in my head into one intense hyper-focused sort of beam […]. It is giving the brain a task that it’s almost designed for. […] give that brain something that really you can tune into and it’s your interest, then all that random stuff just goes boom… I get this incredible intense concentration and that’s great for work….
Artificial Intelligence and Music Ecosystem (2023), is a book that has inspired me during the writing process of this thesis. It discusses what one can achieve by utilizing AI in creative arts. It “highlights the opportunities and rewards associated with the application of AI in the creative arts. […] [The book is] considering the perspectives of developers, students, and artists, as well as the wider themes of law ethics, and philosophy” (Clancy, 2023). The book has been extremely relevant to the thesis. Through it, I have gained a better understanding of how one can work with AI, and its presentation of ethical standpoints has inspired this thesis heavily. There is one chapter that especially stood out to me, called The Artist: Interview with Holly Herndon, “perhaps best known for her experimental electronic music, and for an art practice that spans the art world, academia, and the text industry” (Wiener, 2023).
Like many, Herndon experiments with technology and art. What makes her extremely relevant to the thesis is her view of ethics regarding artificial intelligence. Her views on ethics and AI are comparable to why this thesis is using the CLG, a partly homemade lyric generator (Clancy, 2023, p. 45):
The more I started to think about it, I noticed the ethical implication of scraping up past creative output and then spawning a new version from that source material. It felt problematic to me, and I did not want to use other people’s data without naming them or paying them.
This is the reason, Holly Herndon found a gap, where she only trained her artificial intelligence products with data where she could compensate them (Clancy, 2023, p. 46). This view of essentially stealing others' work feeling wrong, was the initial reason I chose not to use a public AI like Open AI’s ChatGPT and Freshbots’ (2023) AI Song Lyric Generator. These are more powerful and possibly more impressive, however reasoning the huge amount of data the machines have been fed with, it feels ethically wrong. This will be further discussed in Chapter 5, AI and Ethics.
2.2 CLG and Scaler2
“Robotics allows us to explore and achieve new musical possibilities, by combining computer generation with physical sound and embodied agents” (Clancy, 2023, p. 52). A song contains numerous building blocks. This can be lyrics, chord structure, arrangement, production, sound, instrument, or their lack of. This thesis will research how songwriters respond to having artificial intelligence inspire them in this process. The AIa will in this case be used as a tool to trigger lyrics and chord progressions.
Celine’s Lyric Generator, later referred to as CLG, is a lyric generator based on the earlier version, “a machine learning lyric generator trained on 18,000 of my own lyrics”, created by Rachel Meddings. Alongside Meddings, we created a lyric generator based on my back catalog. It contains lyrics from released and unreleased material. It also contains smaller lyrical ideas and unfinished lyrics. The product is limited to, and only fed my own, personal lyrics. The CLG is an artificial intelligence product based on ChatGPT2’s technology. By combining ChatGPT2’s technology and my back catalog, this AI tool is able to produce lines of text. The CLG uses a reference word or sentence as a prefix, which will be the theme of the text produced. This is shown in the video below, with the reference word fake.
Video: Rachel Meddings and Celine Lyng writing a song with the CLG, using the reference word "fake"
The lyric generator produces lyrical ideas so when fed with a search word, such as “fake”, CLG will generate sentences somewhat related to the word. There will be sentences that one finds interesting and wants to write a song around. "Fake, boring, and cold" was one of the lines produced in this video. It is quite a visual sentence, and please note how on-point the CLG is here. I imagine it could be relatively easy to build a story around that sentence, with the chorus ending on that very line. It has used the search word, fake, and by adding three more words, it triggers inspiration.
Figure 1: A screenshot of the CLG's control area
In Figure 1 one can see that the CLG is set to a temperature of 1.5, which one has the option to adjust. When it comes to this AI, temperature means how close to the reference word it is going to produce text. With a low temperature, the CLG will be more restrictive and will most likely have several lines of text that are direct copies of my already written lyrics. An example of this can be seen in Figure 2 below. The reference word to create this text was love, and the AI produced text related to the theme. However, because of the low temperature of 1.5, it created an almost direct copy of one of my already-released songs, Awful Crime. The lines are underlined in the photo. The released lyrics (Lyng, 2021) go like this:
Babe, here I stand accused of my awful crime
I confess forgive me take me back
Babe, here I stand accused of my awful crime
I will never speak to another man
Figure 2: CLG generates a direct copy of already-released lyrics
As the CLG is based on my back catalog, I will recognize if direct copies of my songs occur. This is the privilege of using this AI. Using an AI that is fed a large amount of text for a large amount of authors and songwriters, makes it harder to recognize. However, by increasing the temperature of the CLG, the direct copying of text will happen less often. However, by turning the temperature up, the CLG will also be less limited to the reference word. This will have the AI create text more off-topic. By Rachel and I’s experience, the sweet spot for temperature is around 1,7. It stays on the chosen topic but is still mostly producing new lines of text.
Figure 3: A selection of lines created by the CLG (left) and a finished lyrics (right)
The process Rachel and I use whilst writing goes like this. We open the CLG to trigger inspiration to write a new song. By reading through the text that is generated, we pick out any lines that stand out to us. These lines are copied onto our DAW, which in this case is Logic Pro x. What you see to the left of Figure 3, is a selection of the many lines we found interesting. The sentences that are underlined, are the ones that made it onto the finished lyrics. After structuring these sentences and further developing the idea, the lyrics resulted in what one can see to the right in the photo. The sentences that are underlined here, are the lines that are inspired by what the CLG generated.
By using “a machine learning lyric generator trained on 18,000 of my own lyrics” and the CLG, Meddings and I have experienced a boost in creativity while writing songs together. The spark of creativity can be explained by looking at how the brain connects topics and subjects. If one were to write a song about love, “love” would be the word we put in the search bar of the AI. It would generate text, and we would then have to read through the results and pick out any interesting parts.
However, it does not stop there. “In modern psychological theories about the representation of memory and knowledge, one presumes our knowledge is organized in a semantic network” (Kaufmann, 2015, p.42). This means that if one is given the task of writing a song about the word love, one has personal structures of what “love” is. This can be based on what one knows, has observed, and experiences that have occurred priorly. One’s structure of the word love can look somewhat similar to the figure below.
Figure 4: Example of a semantic network
To create a semantic network regarding the word “love”, I firstly separated it between positive and negative love, and then question how I view, experience or observe positive and negative love. This semantic network is however a simplified version, as for each word that was added, new ideas came to mind. Kaufmann (2015, p.42) explains this such as:
Every time one thinks of something within the network, one will likely get new associations, and the network is drawn longer and longer and becomes more comprehensive. This way it gets closer to another network, and a link can occur
By using the CLG and the search word love, one generates a larger number of text. All referring back to the word love, but in various angles referencing the word love. By this, one is not limited to only what the writer associates with the word love, but their associations to all the words and sentences that are generated through the AI. Therefore, not only will one have a semantic network of the single word love and what one relates to that word. One is left with several, even many semantic networks of all the different ideas that is generated by the CLG. More impressions and associations can result in a bigger number of lyrical ideas, and this can trigger or boost creativity.
It is important to note that the CLG is not capable of generating a fully formed lyric, that will make lyrically or structurally sense. It is not capable of having a discussion about the text generated or generating results by itself. Kaufmann (2019) writes:
The difficulties in fully simulating human thinking in its creative manifestations, is due to the tendency to draw a sharp distinction between cognitive processes and dynamic processes, which are based on needs and emotions. […] It is hard to imagine a mechanical system, like a computer, being able to (by itself), find problems that can be interesting to formulate and pursue.
The CLG requires a human to establish a problem to start with, in this case, it needs a reference word or sentence, to produce text. The generated text also needs to be structured, formed, and adjusted. However, it can be a good starting point for creating lyrical ideas and broadening the semantic network in the process of writing a song.
Scaler2 is the second artificial intelligence product that will be used for this thesis. It is a plugin made by Plugin Boutique, which can be used for creating various musical structures such as melodies, basslines, and chord progressions. “Scaler2 comes with over 400 song and genre-based chord sets and over 200 artist chord sets to inspire new progressions and melodies” (2023). Scaler2 can easily be connected to your preferred Digital Audio Workstation, later referenced as DAW.
For this project, Scaler2 will be used solely to create chord progressions. By selecting an artist or a genre, the AI generates chord structures. The structures that are generated can be altered in various ways. One can also choose to alter the chord progressions, and its key, as well as change the scale one wishes to use. It is possible to change the instrument the chord progression is played in and how it is being played. This helps the writer to hear the structure in various forms. The plugin has many features, such as creating melodies, basslines, and such, but for this thesis, it will solely be used for creating chord progressions. In my experience, this plugin has saved me from repeating chord structures. When limited in one’s accompaniment instrument, one tends to choose similar chord patterns and keys. Scaler2 is trained on a large number of data and will therefore not have this issue.
The thesis belongs to the field of popular music, like the theory that it takes inspiration from. An important focus of the thesis is the ethical standpoint, which also Holly Herndon acknowledges. Consciously, I chose to use a lyric generator that is only trained on my back catalogue, whereas she has gathered content for her AI systems (Clancy, 2023, p. 46). By using the CLG and Scaler2 in session, the participants will gain new experiences. To gather data on their experiences, six interviews will be held.
3 Methodology
To actualize the research questions related to the subject, songwriting, and AI, it is crucial to implement a plan of how to research the topic. One has to make the choice of which methods are efficient. One also has to find a number of participants that fits the thesis. The thesis needs enough participants to gather enough data. However, choosing too many participants can be challenging in terms of time management. The goal of the research is to document how artist react to the use of artificial intelligence in their songwriting process. The thesis aims to document expectations, experience, ethical dilemmas, authenticity, and ownership.
3.1 Context
The Oxford Dictionary describes methodology as “a set of methods and principles used to perform a particular activity”. There is a huge variety of methods one can use to gather data. For this thesis, the main method will be reading articles and books to research the field of music and technology. To collect data on how artists react to the use of AI in their songwriting process, the thesis needs artists who have these experiences. As artificial intelligence is changing rapidly, but still is quite a new public tool, I first had to make sure the artists had had such an experience. To do so, I arranged for one songwriting session with each of the participants, for them to know what writing with AI means. It also let me have participants who had the experience of using the same AI systems. After the sessions, I produced the three demos, which were later sent to the artists, for them to record vocals. To document the artists' opinions regarding the use of AI in songwriting, the artists participated in two interviews each. There was one before the writing session and one after the artists had received a demo version of their songs. Attached to this thesis, one can find the guide for the first interview in Appendix B, and the guide for the final interview in Appendix C.
3.1.1 Research design
To write a thesis around the subject of songwriting and AI, there will be used varies ways to collect data. Firstly, the qualitative research method will be a prominent approach to gathering the right material. The qualitative research method is described by The American Psychological Association (2018), as “a method of research that produces descriptive (non-numerical) data”. The qualitative method that has been used in this thesis, is interviews.
By holding interviews, this thesis will contain information from three different artists, and their experience using AI in their creative process of making lyrics and chord patterns. By holding interviews, one has to choose how one wants to hold the interviews. The interviewer can prepare, structure, and then have the participants answer the questions. One can also hold interviews that are more similar to a conversation, where the questions are not prepared in advance (Jacobsen, 2022, p. 166-168). “Resting in between these two approaches are semi-structured interviews, which follow a predetermined protocol of questions asked of each respondent but allow for spontaneous follow-up questions and variation in how questions are asked (Galletta, 2012)”. For this thesis, the interviews will be semi-structured. Before the interviews, I therefore planned and structured questions, but were not limited to them under the interviews, meaning the different interviews could have quite different outcomes.
Therefore, there will be a set of questions that the participants will most likely be asked. However, the interview will not be limited to or ruled by these questions. A half-structured interview provides more freedom. This format lets the interviewer dig deeper into the answers given by the participants. The conversation can take different turns than expected, which can provide important data. The participants might have opinions or comments that the interviewer had not thought of, and one can then choose to ask more about that subject. However, there will be questions that are prepared beforehand. This is to make sure the most important questions are answered.
This thesis contains interviews with three artists that will be held separately. This is to ensure everyone gets to answer the questions honestly, without their opinions being affected by the other interviewees. When the interviews are held separately, the interviewee does not have to take into account the others and can therefore speak relatively freely about their opinions and perceptions Jacobsen (2010, p. 89). The interviews would ideally be held in person, as it is easier to develop trust and openness (Jacobsen, 2022, p. 165) that way. This was however only possible for one of the interviews, and the others were had over Teams. This brings me to the weakness of holding interviews in person. Even though it is ideal in terms of a conversational flow, it can be expensive in terms of travel costs (Jacobsen, 2022, p. 165). A good second solution was therefore to hold two of the interviews over Teams with the cameras on. Being able to see each other makes it easier to read body language. This can be important, as one wants the participants to feel comfortable at all times. This could be hard to recognize if the interviews were held without the video option.
Three artists were chosen to participate in this thesis, which implies two interviews and one writing session. The artist would participate in one interview before our writing session, where I would gather data on their expectations and prior experience of writing songs with artificial intelligence, and a final interview after receiving a finished demo. In the second interview, the participants were asked about their experience of the session, ownership of the result, and efficiency. It would save time to ask the artist all questions in a single interview, but I chose not to do so, to make sure I gathered the right data. Humans often want to be perceived as consistent in their opinions. That could result in me not picking up on if the participants changed their minds during the process if they were only interviewed after the session.
The interviews were recorded using an online survey tool called Nettskjema, which was developed by the University of Oslo. “[Nettskjema] gives you the option to create, store, and manage surveys and data collection. The service offers […] security measures to ensure data accuracy and privacy” (Nettskjema, 2024). Firstly, I used their app to record the audio of the interviews. They were then saved onto the website, where one easily can access the recorded audio, as well as a transcription of the interviews. The transcription tool saves quite some time, compared to doing the process manually.
The writing sessions held concerning the thesis are meant as a research tool. First, holding three writing sessions where the artists experience working with AI in lyrics and chord structures, gives them more insight to answer my interview questions. The thesis includes one track that represents each artist, as well as the songs written with AI. This allows the readers to make their own opinions regarding how they see the songs fit the artists' catalogs, or not. To do this, we wrote the full lyrics and melodies during the writing sessions. After the session, I would continue to work on the production of the song, as well as recording a lead vocal, for the artists to replace at a later date. When I was pleased with the recording, the songs were sent to the artists for them to record vocals, as well as make changes such as lyrics or musical features. When this was done, I worked on the placement of the vocals and mixing the songs, so they sounded fairly balanced. It is important to note that these songs are demos and not finished recordings. The songs were mainly made and produced for the artists to have something to respond to in the interviews.
As part of Appendix A, I have attached a GANTT Chart, which documents my process of working with the thesis. This includes finding the right artists, arranging and executing the writing sessions, and producing and mixing the songs written in the sessions. The chart also shows the research and writing of the thesis throughout the last 7 months.
3.2 Choosing and introducing the artists
For this thesis, I have been in collaboration with three Norwegian artists. They all have an already developed and established musical sound. The artists are a mix of artists I know, and that I do not. When choosing the artists, the genre was important for me. Therefore, they are all somewhere on the pop-spectre.
It was important for me to find participants that were suitable in several aspects. I needed artists who had developed a clear and distinct sound and/or persona. As the thesis did not allow to use of a bigger number of artists, because of the length of the thesis and the limitation of time, it felt important to have a diverse choice of artists. The artists are in the same genre, but on opposite sides of the pop spectre, from acoustic sounding to electronic music. They also vary in gender, amount of musical education, and number of released songs.
It was important to have artists working in the pop genre.. The reason why genre became important within this task was because I was meant to produce the music we wrote. I therefore needed artists working in a genre I knew how to produce. Therefore, I listened to the artist's previous releases, to ensure I was able to produce a similar-sounding demo for the artist. The songs that are written for the master thesis will be compared with the artist’s earlier releases. Therefore, I need to understand and be able to work in that soundscape. If I was not capable of producing similarly to their already released music, it could also affect the artists' view of the musical results, and even affect their view of using AI in their writing. As I am also a collaborator of the music and lyrics, it also felt right to find artists in the pop genre. It is the umbrella genre my own musical project belongs in, which means I know the writing style but also have a good understanding of the artists' references.
Finding artists with an already-developed sound felt important for this project. For me and them to compare the song we write together with their already written material, only works if they and I know what their sound is. This creates expectations of what the songs we write will sound like, lyrically and musically
The artists will be described and analyzed with their full name/artist name in this thesis. This for it to be possible for readers of the thesis to compare the artists’ already released music with the song we write together.
Introducing Ida Sofie Laland
Ida Sofie Laland is a 25-year-old singer and songwriter from Bryne. Laland and I had our first session on December 5th, 2023. We started with a 15-minute interview where we talked about expectations and how she hears and views her music. She is a part of the indie-pop band, Cherry Cinema, but for this thesis, we will be writing for her acoustic project, Ida Sofie Laland. “This is the type of music where you can play a whole concert using only the guitar”. The project is highly inspired by Hayley Williams’ album FLOWERS for VASES and Taylor Swifts' album, Folklore. “I have felt that the music is touching the Americana genre in a way, though still being indie-pop”, Laland explains.
Before our session, Laland sent me her song Finally Sherry which is written for her acoustic project. This is helpful as her songs are not released and it is easier for me to go into a session knowing what sort of landscape we will be working in. Finally Sherry is a love song, with a distinctly clear vocal. The song is accompanied by an acoustic guitar and some backing vocals. The production also contains some adlibs with some effects and reverb. Going into the session with Laland, I am expecting to be writing a song containing a similar instrumental and arrangement structure. The vocals, lyrics, and guitar will be the main focus of the song.
Introducing Raine St. Frank
Through unforeseen hearing their song, Freaking Me Out, by scrolling on TikTok, I contacted Raine about writing together for this thesis. Raine St. Frank is a 27-year-old singer and songwriter living in Oslo. Signed to the Norwegian label and management, Kjærlig, the artist has released 4 singles. Inspired by artists like Olivia Rodrigo and Billie Eilish, Raine makes catchy electronic music.
I will be using their song, Freaking Me Out, to get to know the artists’ writing and production style before our session. The artist wanted to write something in the same direction as what Billie Eilish and Olivia Rodrigo have released. They want honest lyrics delivered angry sounding, with a big, catchy chorus.
Introducing Torfinn McKenzie
Torfinn McKenzie is the last artist who has contributed to this thesis. Inspired by artists like Bon Iver and Noah Kahn, 25-year-old McKenzie writes strong lyrics, accompanied by thoroughly arranged guitar parts. His sound is a developed combination of genres like folk, singer-songwriter, Americana, and country. To prepare for this session, his already-released song, “Grow Older” was used as a reference. The song has a repetitive and feel-good guitar part. The guitar is what drives the song forward, whilst the tripled vocal’s melodies float upon the steady arrangement.
3.3 Prior experience
None of the participants has utilized artificial intelligence in their songwriting process, and certainly not the way we will be using it for this project. They have not used it to generate chord progressions, nor to produce song lyrics. The three artists are therefore all new to this type of writing. It is therefore crucial to have the three writing sessions, to ensure the artists have experienced using AI tools in their writing.
However, they all have explored AI products such as ChatGPT in various amounts. OpenAI’s ChatGPT, is a «language model [that] can respond to questions and compose various written content, including articles, social media posts, essays, code, and emails” (Hetler, 2023). Torfinn McKenzie (2024) expresses, “I have not used AI to write any lyrics for me. However, I have used it to read through my lyrics and let it interpret them”. McKenzie is therefore using the website to analyze his lyrics, to make sure the story he is conveying makes sense.
The fact that none of the songwriters interviewed for this project has used artificial intelligence to generate lyrics for them, is interesting. It gives the participants somewhat of a similar basis going into the process of writing a song with the help of artificial intelligence.
3.4 Expectations
Artificial Intelligence can be a threatening product to many industries and many are afraid of being replaced across many work fields. This was certainly up for discussion in the spring of 2023 when OpenAI launched the ChatGPT3. Credera published an article (Krishnarai et al., 2023) describing the release of ChatGPT3:
A fascinating model that is free for public use is the GPT3, an advanced text generation model. This AI system takes input text and generates thoughtful continuations of stories or responses to questions using a neural network that is trained to predict the next word in a sequence of words.
This version was more powerful than any AI we had seen before. One could easily ask ChatGPT to write a full song and it can provide a decent lyric in seconds. The lyric generator, CLG is based on the technology of an earlier version, called ChatGPT2.
When I first tried to use artificial intelligence whilst songwriting with Rachel Meddings, I was skeptical. My initial thought was that, if an AI can do this, is my musical education all for granted? If a machine can make a perfect lyric, we do not need the brain to make lyrics anymore. However, the first session working with the tools changed my mind quickly. The AI was not there to do the job I have worked so hard to be able to do, it is there to inspire further writing. It is a tool, not a solution. Because of my rapidly changing mind, I found it important to collect data on what the participants were expecting going into the writing sessions.
4 AI, Lyric and Arrangement
The Cambridge Dictionary (2024) describes artificial intelligence as:
A particular computer system or machine that has some of the qualities that the human brain has, such as the ability to interpret and produce language in a way that seems human, recognize or create images, solve problems and learn from data supplied to it.
The CLG uses the language learnt from my back catalogue and aims to interpret it through the search words supplied by the user. The plugin, Scaler2 has been fed a huge catalogue of musical structures and is therefore interpreting musical structures in various forms. Both of the products have learnt a language and are only able to recreate data similar to what it already has been fed. This limitation is addressed by Harhaug (2020):
No matter how many parameters and possibilities we give the machine (limited only by the technical abilities of how much it can process, which is going to at least double every year from now on) it’s still a machine that will only do what we tell it to do. It can only imitate and copy.
Though imitating human intelligence, it still only knows what it is programmed to know, and it though similar, is not a living, breathing creation yet. Artificial General Intelligence, or AGI is at this point only a theoretical but can be viewed as a continuation of the AIs we use today. "Someday, AGI may replicate human-like cognitive abilities including reasoning, problem-solving, perception, learning, and language comprehension" (McKinsey & Company, 2024). However, this AI system is still decades, if not centuries, away, (McKinsey & Company, 2024) and this research paper will therefore focus on where AI is today.
"As humans, we so often feel helpless in our smallness, yet still we find the resilience to do and make beautiful things, and this is where the meaning of life resides", Nick Cave (2023) writes. Humans find joy, and maybe even a higher meaning, through creating art. Why would we then use a robot to be more efficient, if the creation is the goal itself? For me, the answer lies in the choice of which AI one uses. Choosing an AI that creates music for one, is not necessarily rewarding. However, using an AI to trigger inspiration, resulting in further human creation, can be rewarding.
For this thesis, the research is solely based on the AI systems, CLG and Scaler2. For a further, in-depth understanding of what artificial intelligence is, please continue the research through The Cambridge Handbook of Artificial Intelligence (2022).
4.1 AI and ethics
Before using artificial intelligence, one should acknowledge and consider the ethical implications of using the products. This chapter will present ethical dilemmas regarding AI and the testing of two major AIs. There are several issues regarding artificial intelligence and ethics, such as the concerns of copyright and consent. The copyright laws will grant ownership to the songwriters who have contributed to writing the words and/or arrangement of the song. However, as Botpress puts it (2023), “AI blurs the line between human intervention and machine creation, reasoning the copyright infringement”. To understand why copyright can be an issue in regards to using artificial intelligence, one has to ask, what are the large language models trained on and who owns the results produced by the AIs?
“The largest AI’s we see today are virtually trained on all text material that is on the internet” Mike Priest (2024) explains, clarifying how artificial intelligence works. This text material is text data such as books, articles, or web pages (Priest, 2024). It is important to note that the people producing this text were not given the chance to give their permission for their work to be used. It was simply done without consent. This is something Bruce Houghton (2023) finds wrong, writing:
The AI engines are violating the rights of the authors and creators of the underlying materials and even if the result is not an exact copy of the underlying work (think “sampling” in music), the resulting AI product would not exist “but for” the underlying work and the use of the AI work competes with the artists or creators’ ability to make a living.
To have AI create lyrics can from only that, feel ethically wrong. To know that their hard work has been taken and fed through a machine. Choosing to use the AIs despite the issue of consent, brings up another issue, which is addressed in the book, Artificially Intelligence and Music Ecosystem (Clancy, 2023, p. 54), writing, “one of the main open questions in the regard is who owns the product of creative AI system – the dataset creators, the system designers, the public, or maybe the machine itself?” Not only are the ones that created the data not credited in the biggest AIs, but as artificial intelligence is evolving so rapidly, laws around it are not in place, and we do not know who owns the results generated.
The copyright infringement refers to “the unauthorized use or reproduction of someone else’s copyrighted work” (Botpress, 2023). This means that by using artificial intelligence, like Botpress’ lyric generator or OpenAI’s ChatGPT for writing song lyrics, one is in danger of using someone else’s text. Botpress (2023) continues stating “[…] it may raise legal concerns about potential violations of copyright law”. With artificial intelligence and law around it evolving rapidly, using the tool in songwriting and releasing the material seems risky. Freshbots (2023) has created an AI Song Lyric Generator which is an artificial intelligence product that generates song lyrics. By using AIs like these, one should consider if it is safe to use the results in terms of copyright. This is a point Freshbot (2024) themselves brings up:
While we are striving towards a future where duplication and similarities are completely non-existent, due to the model's training on internet lyrics, some generated content may bear similarities to existing ones. […] We strongly recommend conducting a thorough plagiarism check before proceeding, to avoid any potential copyright issues.
I had the Freshbot's AI write me a song about being lonely, which also ChatGPT and the CLG will be tested further down in this chapter. By using the AI Song Lyric Generator, one is asked to select some keywords, emotions, topic ideas, and/or genre, and the generator will create a song lyric tailored for the users in a few seconds.
Figure 5: Freshbots’ (2023) AI Song Lyric Generator’s lyrics
about being lonely
The songs are structured into parts such as verse, chorus, bridge, and outro. The lyrics contain rhymes, which are commonly used in song lyrics and metaphors to paint a picture for the reader/listener. This story told through the lyrics makes sense and it looks much like a song. However, it is not perfect.
Explaining how she creates music, Kari Iveland (2024) writes that it "involves a simultaneous act of listening and singing, engaging body and mind in ways that often bypass the awareness of a conscious mind". Songwriting is a connection with the words, the music, and the body. Having an AI write the full lyric interrupts this thought. Looking at the lyrics created by Freshbots’ (2023) AI Song Lyric Generator’, the connection between the words, music, and body is not in balance.
Though not all styles of music are strict regarding syllables and repetition of melodies, the genre of pop often is, and it was the first thing I thought this lyric was lacking. While trying to make a melody for these lyrics, I found it hard to create melody lines that fit both verses. A melody could work for one, and not the other. However, it seems as if the generator is trained to count syllables. The amount of syllables used for each line in the first verse is also used in verse two. This is common in pop songs, but the AI's lyrics still do not fully work, unless you make melodic or rhythmical changes. One way of singing it is in 4/4, with two bars per line, as shown in the figure below, with a rhythm of 8 eight-notes, followed by a whole note at the end.
Figure 6: A visualization of how lyrics lines, with the same amount of syllables do not necessarily work
Verse one works perfectly, with an equal length for each syllable, and a long note at the end, where the word “go” rings for a longer amount of time. The melody and rhythm of the melody work well. Iveland (2024) writes:
Lyrics and voices [...] provide more than meaning or holders of sound, offering rhythmical and prosodic qualities and a story and sounds for the singer to convey.
Repeating the rhythm for verse two, however, does not work completely. The way the word “go” rings, the word “shatter” does not, reasoning it ending on a vowel. Firstly, it has two syllables instead of one, but the problem also lies in the placement of the rhythm. The beat falls on the last syllable of the word, which also creates a problem. Shat-ter would be the correct way to say it, whilst in this lyric, following the same rhythm of verse one, would result in the vocalist singing shat-ter. Chomsky (2023) writes:
This is a good example of AI's needing humans to alter what is generated. The AI can produce songs for chosen genres and moods, but it is only as smart as its data. As we saw in the example above, the AI can rhyme and use the same amount of syllables in verses. However, it is not able to make the lyrics that are easily sung, as shown in Figure 4, at this point.
An AI that could have been natural to use for lyric writing for this thesis, was OpenAI’s ChatGPT, due to it being used by so many. Bernard Marr (2023) has described the evolution of ChatGPT and a short extract can be read below:
ChatGPT has had a profound influence on the evolution of AI, paving the way for advancements in natural language understanding and generation. It has demonstrated the effectiveness of transformer-based models for language tasks, which has encouraged other AI researchers to adopt and refine this architecture.
The figure below shows the difference between how the CLG and ChatGPT produce lyrics. This test is done by asking the AIs to write lyrics about being lonely. To the left is a part of what the CLG produced. Many lines fit well with the subject, which can trigger further inspiration. However, the CLG is not capable of deliberate rhyming, structuring, or being consistent with one idea throughout, other than keeping the lyrics in the first person.
Figure 7: CLG (left) and ChatGPT’s (2024) (right) songs about being lonely
ChatGPT can effortlessly generate lyrics, as it is fed a large number of information, compared to the CLG which is only fed my back catalog. ChatGPT produces better-formulated ideas and one can communicate with it on a much higher level. The lyric is structured in a pop format, which refers to the structure of a song. “The pop song structure can take varying forms but will typically involve a verse/chorus/verse/chorus/bridge/chorus structure” (Hope, 2023). Pop songs likely contain both rhyme and/or near rhymes. The lyric created through ChatGPT is consistently on theme and can use metaphors. An example of a metaphor made by the AI is “echoes of laughter in a silent home”, which can refer to a house that once was filled with people and happiness, and now is silent, and one is only left with the memories.
What ChatGPT (2024) produced, compared to the CLG, looks impressive independently, but seems to have the same issue as the AI Song Lyric Generator. It lacks structured sentences in terms of the amount of syllables. In fact, none of the sentences in the two verses hold the same amount. One could say that the two AIs hold different purposes. As someone not used to writing lyrics, ChatGPT can be an amazing tool. It can provide a decent lyric with very little detail, as shown in Figure 5. However, as a songwriter, it can be more inspiring to be using the CLG.
ChatGPT is a powerful AI and can be used for various tasks. “About three-in-ten employed adults under 30 (31%) say they have used it for tasks at work”. Reading the thesis might therefore be more relatable to the reader if ChatGPT was used for producing lyrics. The CLG, on the other hand, can only be used by Meddings and I for now. However, I did choose to use the CLG. This is partly because of the copyright and ethical implications that come with using AIs such as OpenAI's ChatGPT, but not only that.
An important reason why I chose to use the CLG was its limitation. "The human mind is a surprisingly efficient and even elegant system that operates with small amounts of information" (Chomsky, 2023). As it is only fed with my back catalog, the AI is limited, and this happens to generate more focused results. Through what the AI generates, the mind sufficiently connects the results and inspires for further writing. This point, in addition to the risk of copyright infringement and the lack of consent, makes large language models less attractive to me.
The artificial intelligence plugin, Scaler2 is also fed with a lot of genres and artists to generate musical structures. This can also challenge the ethical view. However, as we are only using it to create chord progressions, which are not copyrighted, I do not see any implications for using the tool in the sessions.
Lastly, if one feels strongly about the subject of one’s art being trained to make AI products smarter, there can be ways out. Jess Weatherbred (2024) wrote an article in The Verge about different ways of doing this, focusing on image-based art. However, she also acknowledges that:
Creatives in other fields, like writing, voice acting, and music, are also fighting to protect their work. It’s much harder to disrupt how AI models are trained on this kind of data without noticeably affecting the original content.
She encourages musicians to read and understand the user terms of any hosting platform where one is uploading musical works. If publishing the music via a personal website, one can block GPTBot (Weatherbred, 2024).
4.2 Sessions
To prepare for the interviews and the writing sessions, it was important for me to be comfortable to hold the meetings over Teams. To use the CLG is a process. To load the AI takes about 45 minutes, as one is teaching the AI the language, which in this case is my back catalog. To do so, I had a session with Rachel Meddings, trying out and exploring the CLG together. It was also important for me that the interview- and writing sessions ran smoothly. Therefore, I had a practice run together with my sister, Lene Lyng, where we tested sound and visials. It was important for me that I could see the participants during the interview and that they could see me. As for the writing session, there were other priorities. Though keeping the cameras on, seeing each other was not the most important part. It was important for them to see the CLG and the Logic project. To see the CLG is important whilst working on the lyric, as we were picking out and discussing the text that was generated. The lines of text we liked from the CLG’s results were then put into the text option of Logic. For the participants to be able to see the Logic project was also important, as it is where we structured and finished the lyrics. As we were not in the same room, it felt important that they could see what was happening on the computer.
Sessions stagnating because of technical difficulties can create stress for me, as I was the one in charge of the technology. This then again affects creativity, which has been mentioned before. It might also affect the participants waiting for things to work too, in terms of time, and stress levels and it being simply boring to wait for technology to work. This was the reason I made sure I had control on my part. I knew how to control what the participants saw and heard during the session.
The sessions with the three artists were structured in the same way. We would first have a few minutes talking about unrelated topics. This is a technique of making sure we feel comfortable and make the situation less stressful. This was followed by a short interview about their expectations to be using AI tools in the songwriting session. This lasted between three and a half minutes and up to almost 13 minutes. After the interviews, the writing session would start.
Psychological safety is important for the dynamics of the session. “Psychological safety is about condor, about making it possible for productive disagreement and free exchange of ideas” (Edmondson, 2019, p. 15-16). It is important to gain psychological safety and to feel welcome to share ideas openly. Naturally, all ideas cannot be used, but by creating a comfortable space, the participants can communicate ideas, provide input, and explore different ideas freely, without the worry of rejection.
Songwriting can feel revealing and personal for many, as one often writes about one’s life and experiences, good and bad. One needs to create a welcoming environment so one feels comfortable coming up with ideas. Therefore, it is important to be aware of one’s communication. “The first requirement to good communication is to show interest. One can show this by the way one listens. The ability to not speak, but listen, is actually crucial for good communication with others” (Gjøsund & Huseby, p. 93). This is important in all social situations, and it also applies to songwriting sessions. It is essential to take the participants' ideas seriously and consider them.
The structure of the sessions would be to start with the CLG, where we produced lyrics in collaboration with the lyric generator. When we had a clear idea of what the song would be about and were satisfied with the amount of lines of text and ideas we got from the lyric generator, we opened Logic Pro X. This is the DAW the songs were produced in. We proceeded with opening the plugin, Scaler2 to work on a chord pattern. When this was done, it was all up to the artist and me. We would use our lasting time of the session to structure the lyrics into parts like verses and choruses, make melody lines, and talk about where we could see the song going.
After this, we ended the session and I continued producing the song, as described in the session and referencing their back catalog. I recorded a guide vocal and then sent the project to the artists, for them to record lead vocals and add any ideas they had. These ideas can be lyrical, melodic, and for production. This type of collaboration is called asynchronicity (Bennett, 2011), where process of writing the song is kept separately and with no defined roles. Personally, I believe it is natural for the participants to have new ideas when hearing the demo. It is only natural to see the songs in a new light when coming back to the song after some weeks.
The songs were written in the span of 2-4 hours.
4.3 The music
As part of the research for this thesis, there has been written one song with each participant. The songs were written for their musical projects which means it should sound somewhat similar to the rest of their catalogue. Therefore, as mentioned when introducing the artists, I chose a song from their portfolio as a reference to production and sound. There are many relevant questions when discussing the subject of fitting into their catalog. With using AI, will it lack any human aspect? Will the artist relate to the song in the same way as their other material? Will they feel the same ownership to this song, compared to their other songs? Will the AI trigger inspiration or will it be distracting?
4.3.1 Ida Sofie Laland
The first session was with singer and songwriter, Ida Sofie Laland and was held over Zoom. After the first interview, we started the writing session, starting with the CLG. The reference I used to prepare for the session is her song called Finally Sherry. The song has also been referenced in the production of the song Laland and I wrote together.
Audio 1: Finally Sherry by Ida Sofie Laland
Figure 8: Ida Sofie Laland and Celine Lyng working on the song “Cold”
The picture above was taken in our songwriting session. This was held over Teams, where we used the CLG for lyrics and Scaler2 for chord progressions. To the right of the picture one can see Laland and I, and on the left one can see the song Cold being produced in Logic Pro X.
Ida came to the session with the idea of writing a song that fit the season. At the time of our session, it was cold and snowing outside. We therefore portrayed the scenery of someone who is longing to go to their friend but struggles to do so because of the cold. This coldness we write about is however not referencing only the weather, but a yearning for a friendship that once was.
Triggered the CLG by choosing words related to the theme, such as cold. We thereafter spent some time reading through the text that was generated. Any lines or words that stuck out to us were copied and pasted into the Logic project. A selection of such lines can be viewed in the photo below:
Figure 9: A selection of the sentences generated by CLG that we used for further inspiration and writing
At the time of the screenshot being taken, the sentences in Figure 9 had not been structured. This means that the lines are directly copied from the CLG and not put into context of song structure. Some of the lines were used in the final lyrics, some were edited and some did not make it.
The second to last line of Figure 9 is, They say that memories keep you company. The line resonated with Laland and me and became the start of the chorus. Other lines that were used had to be restructured. The lyrics of verse two ended up like this:
I can't help but feeling kinda silly
Maybe I am immature
I could stand outside your wooden door
but can I make it in the cold
The first two sentences are almost direct copies of what the CLG produced. The lines from the CLG were originally I feel kind of silly and maybe I am immature. Some lines were also not used in the final lyrics.
The song's arrangement is the same as the reference song, Finally Sherry, with a verse, chorus, verse, chorus, bridge, and a final chorus. This is a standard arrangement in the pop genre, meaning the structure of verse, chorus, verse, chorus, bridge, chorus, which is suiting. Like Finally Sherry, our AI song, Cold, features a lead vocal, guitar, vocal harmonies, and some bass.
Up to that point of a finished topline and chord progression, we had made all the decisions together. After our session, I prepared a demo of the song which contained a guitar and a lead vocal. She, like the others was then to re-record these vocals. All were allowed and encouraged to change any lyric or vocal line if they had any new ideas, as well as to come up with production ideas. Laland made some lyrical changes to enhance the story but still kept the wording. The lyrics to Cold can be found in Appendix D.
Audio 2: Cold
Ida Sofie Laland’s vocals are the lead in both songs and her vocal delivery is similar. Laland has a clear voice with stable pitching. Her high-pitched vocals attack the notes head-on, contrary to gliding between notes. This makes the melodies clear and her voice recognizable.
The same guitarist was used for the recording of Finally Sherry and Cold. This enhances a similar sound and arrangement of the guitar in the songs. The acoustic guitar sound is warm and round, meaning the guitar produces sound with dominating mid frequencies and with the high frequencies lowered, simulating a tube.
4.3.2 Raine St. Frank
The second session held for this project was with Raine St. Frank. Raine expressed a wish for a song with a big chorus and an angry expression. To prepare for the session and production purposes, their song “Freaking Me Out” was used.
Audio 3: Freaking Me Out
Figure 10: A screenshot of Plugin Boutique’s Scaler2.
The picture above shows what the Scaler2 produced in the session with Raine St. Frank. The chord structure shown below the keyboard was the initial inspiration for Raine and I’s song The Things I Know. Below the chords, one has the option to choose which scale one wishes the song to have, which in this picture shows a Lydian mode. Further below it gives some theory to the scale, as it shows which chords one finds in the set scale. Farthest down, there are eight empty rectangles, which gives one the option to alter the chord structures manually, by dragging the chords into one of the rectangles. This is helpful for instance when one likes the chord structure given but wants to switch out a couple of chords. When one is happy with a chord structure one can easily drag the chord progressions to the project, and it will appear as a MIDI file.
Together with the help of AI, we wrote “The Things I Know”, telling the story of struggling and how entitled people can be. They keep telling one how to feel, what to do, and how to act when one already knows or has tried these things. Referencing their earlier released song, “Freaking Me Out”, the production is heavily electronic and has many vocal layers.
This was the song I found the most challenging to produce. The production of Raine St. Frank’s songs is bigger, in terms of production layers than what I am used to. This was therefore the song that took the longest to produce but was a good challenge. The finished demo can be heard below and the lyrics are attached in Appendix E.
Audio 4: The Things I Know
4.3.3 Torfinn McKenzie
The last songwriting session that was held for this thesis was with Torfinn McKenzie, and the reference that was used was his first release, Grow Older. This was the only session for the thesis that happened in person, whereas the other songs were made through Teams. Using Teams for sessions is a great option when one lives far from the other. However, it makes it harder to sing and play together, as it usually causes issues with the audio.
Audio 5: Growing Old
Figure 11: Torfinn McKenzie and Celine Lyng working on the song “Memories Strife”
The session with Torfinn McKenzie was held in my apartment. We later met up once more to record some more guitar tracks in his apartment, which is where the picture above was taken. As Torfinn McKenzie and I were able to meet up, it was easier to play and sing together more during the session, which affected the music.
After writing the lyrics and chord structures, we sang the song together several times, where we made small changes to the melody, structure, and wording. This is one of the limitations of using Teams, and the strength of writing in the same room. Personally, I feel like the song becomes more alive when working in the same room. For me, when going straight to production, and without the step of playing, it feels harder to figure out where one wants the song to go, arrangement-wise and dynamically.
Writing songs in the same room opens up for jamming, something Joe Bennett (2011) describes as a process where “a band creates live ideas in the rehearsal room, forming the song from individual contributions to the arrangement and some degree of veto (e.g. U2. Band members may bring stimuli to the session (titles, riffs, etc)”. Though using the same process, using the CLG and Scaler2 at the start of the session, we ended with jamming the song. This opened up for us both to explore ideas on the software, but also to use instruments and experiment with melodies, chord structures, and riffs and in a way, be more playful. The guitar riff was further developed and made us question if the song needed keys. While singing together we perfected the melodies and it was also the reason the song ended as a duet. The song can be heard below, and the lyrics can be found in Appendix F.
Audio 6: Memories Strife
5 Analysis
Through the analysis of this thesis, one will find answers to the research question, as well as the other questions mentioned in the introduction of the thesis. The analysis will also contain some of my opinions regarding the songs that were written in correlation to this thesis. The reasoning is the fact that the creation of these songs was done as a collaboration, and to have several views to the question concerning if the songs fit their back catalog.
5.1 Expectations vs Experience
During this project, the participants went through two sets of interviews. Firstly, a brief interview where we spoke about their prior experiences of using AI and their expectations for using it in their musical writing. The second interview that was held after the artists had received a demo of the songs they had written with AI. The last interview was longer than the first one, where we went more in-depth about the experiences using the tools and their thoughts and more generally around the subject.
The artists’ expectations for using the CLG and Scaler2 in the sessions were somewhat divided. Raine St. Frank explained in their first interview that they were feeling sceptical to be using AI in their songwriting, but were open to try and believed they could be useful tools. Raine adds that they were positively surprised by the AI’s:
I do not know what to tell you, I was surprised. I was left with a positive feeling and a positive experience following the session. The feeling of still having control of where one wants to take the lyrics but getting word starters or sentences starters. It was a very nice tool to gather inspiration, words and sentences.
By word starters and sentence starters, they are referencing the ideas that were created with the CLG. The AI-generated words and sentences, would not necessarily be used how the AI initially wrote them, but it creates a starting point.
In the first interview, before our session Ida Sofie Laland explains, “I definitely have expectations to be using AI, but feel like I have found my method”. Within her method, she explains how she often finds inspiration from chord progressions of songs she likes. Laland had not tried Scaler2 beforehand, but the two techniques are not very different. Scaler2 is fed with a huge catalogue of songs and will therefore generate chord progressions used in well-known songs. Another thing she brings up is an expectation that the process of writing with AI to be quicker than writing without it, which she did not necessarily view as a good thing. I was therefore interested to hear what he had to say about this, after the writing session. Laland explains that the CLG gave the writing session some sort of kick-start:
It works by giving suggestions, and even if the text doesn't end up being exactly how it was generated, it is not certain that the idea would have come if we had not had the inspiration in the first place.
The lyrics Laland and I wrote, were like the other sessions, inspired by the results from the CLG. Though re-written, it is like Laland says, it triggered an idea.
Personally, I found it interesting to hear guitarist McKenzie’s view on relying on AI to produce chord progressions. Chord progressions are fundamental in his music, and it could be expected for him to feel sceptical about such a tool. McKenzie would normally start with the guitar, making patterns and melody themes. As the arrangement of the guitar is such a big part of his music, it could feel restricting or uninspiring for him. When asking McKenzie about using Scaler2 for generating chord patterns, mixed feeling towards the AI was shared. As for the negative sides, McKenzie recognises just this, he is a guitarist, first and foremost. “My songs dig deep into the arrangement. I like the freedom of using the guitar and find it easier to test out things, this way”.
Personally, I understand why McKenzie might feel this way about Scaler2 and be somewhat prepared for his reaction. It is the same reaction I had when researching for this thesis and asked both ChatGPT and A Song Lyric Generator to write me a song about loneliness. The lyric is written in seconds and is fully structured, containing rhythms, metaphors and so on. To use these AI’s as tools, one would have to take something that makes somewhat sense independently and make something new. Creativity usually relates to solving a problem, but by using these highly capable AIs, there is no problem to solve, only smaller details. In the way McKenzie wants more from the arrangement, and I in a way, want less from the lyric generator. we want to feel inspired but to solve the puzzle ourselves. McKenzie wants to create the arrangement by himself and his instruments.
Another valid point McKenzie recognizes for the generator is its limited options for genres. “It did not have a specific genre that suited my style, so a compromise was made from the get-go”. Though Scaler2 has been fed with a huge amount of music, it did not have a genre that suited McKenzie. None of the genres McKenzie recognizes himself in, such as Americana, singer/songwriter, country or indie-pop, and were therefore having to compromise by using the genre of pop balled. This is something also Meddings and I have recognized and compromised on. Our music is in the genre of indie-pop and bedroom-pop. However, as they are not options on Scaler2, we are forced into a standard pop genre.
Figure 12: The different genres one can utilize using Scaler2 by Plugin Boutique (2024).
However, there were not only negative opinions to say about Plugin Boutique’s Scaler2. McKenzie explains that it is a good tool to get an initiate feeling or mood. “You get a sense of mood from the chord progression. From that, I get ideas. I liked how it could give feelings and impressions and a vibe”. To work with the AI in this way is similar to how we used the CLG. Using the tools this way, one strictly uses the AI for inspiration and triggers for further development of the song.
5.2 Efficient or distracting
To be working with a new tool like AI can be inspiring and exciting for some and distracting and “in the way” for others. It was therefore interesting to hear what the participants of the thesis ‘opinions were. The tools are there to trigger ideas, lyrically, melodically, and structurally, by generating text and chord patterns. However, it could also feel distracting to be using new tools, when one has been making music without the tools for so long in the past.
Something Laland confessed in her first interview was her worry about efficiency. Using AI to help write songs both on the lyrical and musical aspects, can make the process quicker. Laland expresses that she has divided opinions on efficiency and that it can be hard for her if things are moving too quickly. She goes on to say that the process also can be quick without feeling uncomfortable.
Torfinn McKenzie finds AI efficient in creating new ideas. “It does not stagnate the process, because of the way we used the AI’s. They were used as tools and a stepping stone for finding new associations and ideas”. The AI was therefore effective in triggering lyrical ideas and melodic and chordal ideas, according to McKenzie. However, he adds that “if we were to use only what the CLG and Scaler2 produced, the process would likely stagnate”. This is only natural, especially for the CLG as it is not smart enough to produce a full text that works on its own. It has to be structured, often rewritten and most of it will not be used. However, Scaler2 can generate chord structures which it is possible to use throughout the whole song. Even though it is able, it does not mean one prefers to use it like that. McKenzie prefers writing the arrangement organically with a guitar or piano.
5.3 The feeling of ownership
The question regarding whether the artists feel the same type of ownership to the songs written with artificial intelligence, compared to their other material, feels important for the thesis. For the AI tools to be considered productive and effective, the product must not compromise the artists' feeling of ownership. They should not be used at the expense of it. The authentic part of the music is likely coming from the artist and not the AI, and it is therefore the question considered important.
In Raine St. Frank’s first interview, they confessed a wonder about the final product, and if it would feel like theirs or not. There has to be some persistence in the music, persona, and image, for artists to be recognized as authentic. Bloomfield (1993, p. 16-17) writes:
In the mass consumption of music, pop songs at their most effective provide the listener with the illusion of entering into a direct and immediate (unmediated) relation with the human producer that is capable of gratifying the listener’s individual need, that speaks directly from one subjectivity to another.
Authenticity is important for performing artists and their audience, and it was important for me to do my part to help the situation production-wise. This is the reason for me to use reference tracks from and for each artist. However, as we are using AI, I would not have any control over what the AI would produce and it was up to the artists and me, to decide how much we would let the AI affect the resulting lyrics and chord progressions.
In Raine’s second interview, they were asked about authenticity and the song, The Things I Know. The artists say as follows, “No, I do not believe it affected the authenticity, as we only used the AI’s for inspiration. We did use some of the lyrical lines given to us by the AI, but we built it further with our human abilities”. Savery and Weinberg write (Clancy, 2023, p. 53), “human collaborators, on their part, bring their unique human advantages to each interaction, such as emotion, expressivity and creativity”. These qualities are what makes the artist to be seen as authentic, or not. Their emotion, expressivity, and creativity with the music are also what can make the artist experience the song to be authentic to them. That means that even though Raine felt slightly skeptical of whether the result of the session would lack authenticity, that was not the case for this session. The artists add that they feel the same type of ownership to The Things I Know as their other songs.
Nielsen (1994, p.160) believes that a piece of music constitutes a whole universe of meaning, a big specter of experiential possibilities. Music contains acoustic layers, structural layers, bodily (motoric) layers, tension layers, emotional layers, and spiritual layers. Viewing music in this sense makes it clear that AI, on its own, cannot produce music or lyrics alone. The lyric generator, for example, does not connect with emotion. It only knows the text that it has been trained on, and even though it imitates how humans work, it will not have the power to understand emotions.
However, humen are inspired by many things whilst creating, and that does not necessarily damage the artists authenticity. By a human being inspired by AI, and thereafter restructuring and reworking the generated material, it does not necessarily hurt the artists' experience of what is authentic to them. Torfinn McKenzie say as such:
I showed our songs to someone, and I told them I wrote this with Celine. And for me it did not feel that the song was written with Celine and an AI. When I hear the song, I feel we have 100% ownership to it, because the inspiration we got from the AI, could have happened without the tools, by spending more time on the song. It just shortens the process.
McKenzie’s experience of the use of artificial intelligence is similar to Raine St. Frank's perception, where they both feel ownership of the song, uninterrupted by the AI. When it comes to Ida Sofie Laland, new thoughts were brought up. Earlier in the thesis, I mentioned how a session could paint the artists' picture of the AIs. This could be because of unrelated factors.
5.4 AI song vs back catalogue
When asked to participate in the research of this thesis, the artists were made aware that we were going to be writing for their musical projects. The initial goal is therefore that the result of the sessions are songs that fit straight into their repertoire. However, there is the question if the use of artificial intelligence would create any distance between the artist and their AI song. Will the song lack some human aspect? Will the songs feel less authentic to the artists?
Raine St. Frank is in the process of creating their first album which is planned to be released in 2025, and they express that The Things I Know will most likely be a part of it. “The album will be in a similar vibe to what we made. It will be a mix of sad and angry songs, about depression and anxiety. The song fits perfectly”. The conveying of emotions is important for the artists. Ruud (1997, p. 81) writes “The music is used to describe, translate and clarify the inner movements we call emotions”. Raine is a good example of this. For our session, they already had a feeling in mind, and it was the initial idea for the song. This idea was further developed in collaboration with myself and the use of artificial intelligence. Though both the AI and I were new elements in their writing session, the artist did not experience that it created any distance between them and their art and the result felt authentic enough for it to be used for a later release. This indicates that though AI was a big part of the process of writing the song, it did not affect the authenticity of the result.
Both Raine St. Frank and Torfinn McKenzie agreed with their feelings towards the artificial intelligence-inspired songs. They saw no reason for the song to lack authenticity, reasoning the way the AI was used. McKenzie sees the reason for the AI song to feel authentic, has to do with how the tools were used. He said it worked “because of the way we used it. It was solely used as a tool and a stepping stone to find associations and ideas”. This way of working with artificial intelligence is different than for example the lyrics produced by Freshbots’ AI Song Lyric Generator (viewed in figure 5) or OpenAI’s ChatGPT (figure 7, right side), which has been mentioned before. Using the CLG will not produce a finished lyric, but can trigger ideas and give text, associations, and sentences. This appears more inspiring than having a full lyric provided from a one-word search like one can with the other lyric generators mentioned above. For the artists to feel that their songs are human-made and fit their catalog, they need to feel a part of the writing process to, then again feel ownership of the result.
It is interesting to me that the artists see their songs as not disturbed by the use of artificial intelligence. They express they feel the same type of ownership of the songs, uninterrupted by the machine that inspired both the chords and lyrics. Though artificial intelligence was taking the lead when it came to lyrics and chords, it had little to no impact on the structure of the songs. It might be because the structure of songs has become part of one as a writer. Raine and Laland keep their songs in a standard pop structure, and whilst writing they already structuring the lyrics into verses and choruses. McKenzie However, who is not following the pop structure, chooses to do so intentionally. Timothy Warner (2020) writes:
It is possible to view the standard pop song structure as a process of expectation and resolution: introductions, verses, bridges and middle-eights provide a sense of expectation in the listener that is only resolved by the chorus.
Comparing Ida Sofie Lalands's Finally Sherry and Cold makes it obvious that the structure and length of the song are extremely similar. Laland uses a pop structure consistent with verses, choruses, and bridges. As shown in the figure below, the only difference between the structure of the songs is that Finally Sherry contains an intro. The songs are also similar in length, with only 10 seconds separating them. Comparing the different parts also points out that each section is, give or take, 30 seconds.
Figure 13: A comparison of the structure of Cold and Finally Sherry.
In the second interview with Torfinn McKenzie, he made it clear that structure, or the lack of it, was an important part of his music. Taking inspiration from artists like Bon Iver, McKenzie expresses:
A characteristic of mine are an untraditional way of building my songs. For example, my song Grow Older is structured as verse/pre-chorus/verse/pre-chorus/chorus and then it is done. It does not matter much what people think, in my opinion
Similar to what was mentioned regarding the unaffected structure of Laland's songs, also appears in writing with McKenzie. Like McKenzie’s song Grow Older, also our song, Memories Strife lack of traditional structuring and clear separation between each part. The songs do not contain classic choruses, which are often looked at as the most important part of a song. Timothy Warner (2020) states that “many pop songs save the most memorable and catchy material for the chorus, which often functions as the ‘hook’ – the element that most listeners find most interesting and fulfilling”.
Whilst writing Memories Strife, McKenzie and I never discussed the structure of the song with names such as verses and choruses. The song was created by structuring the melodies and lyrical lines, followed by singing the song over and over together. The figure below shows the structure of Memories Strife and Growing Older.
Figure 14: The structure of Memories Strife and Growing Older.
In the interview, McKenzie says, “The songs do not have a chorus, but in a way, the pre-chorus [part C] feels like a chorus the second time around. It gives the part a new meaning”. McKenzie therefore, in a way views the C-part as a type of chorus. I, on the other hand, view the A-part, the instrumental, as the main part of the chorus. As we did not discuss this whilst writing the song, we had different experiences of how we viewed the parts.
I do not believe that the fact that we do not view the parts alike, has a negative impact. […] All the parts are there, and they work. […] It just breathes, it just is. And if one does it with quality, like we do, it will be good.
In that way, Warner’s comment connects our different opinions on what the chorus of Memories Strife could be, as it simply is the part that we found the most interesting. For McKenzie, that is the B-part and for me, it is the A-part.
6 Discussion
For the discussion part of this thesis, the sub-research questions, mentioned in the introduction, will be answered and reflected upon through the data gathered from the interviews with the participants. The discussion involves ownership and efficiency, and how they see the songs written with AI fit the rest of their catalogue.
Will the songs made with AI lose any human aspect or hinder authenticity?
"The music is compared to a mirror that reflects strong signals to other social groups about who the users of the music are" (Ruud, 1997, p. 107). In this case, the music does not only reflect who the users are but also how it affects the music the artists themselves make. Raine St. Frank is a great example of this. The emotions they portray through lyrics, accompanied by their strong, vocals, feels believable and for me perceived as authentic. “Even though it feels hard to define oneself, I feel that it is part of my artist identity. The melancholy world, in a way, or a type of melancholy universe”. The sad and the angry, which is the theme of The Things I Know and their upcoming album, fits perfectly in that universe. This might be the reason the songs written for this thesis fit with their authenticity, according to them. Emotions are important for the artist and were the focal point throughout the writing session and whilst producing.
Do the artists see the newly written songs fit with their back catalogue?
As mentioned, Raine St. Frank saw her AI song fit their back catalogue and is planning to release it as part of an upcoming album. This is an indication that songs written with and without AI tools can fit together as part of the same catalogue, at least for them. Torfinn McKenzie also believes Memories Strife has a place in his repertoire, and states the song fits his catalogue. However, as he only recently has started to release music, he adds:
I do not see why it would not fit, but I do not think it fits with the first releases. However, I can easily see it fitting great to the next. It feels as if the song is a further development of me as an artist.
Personally, I agree with the view of deciding to wait to release a duet to his audience. When one is starting to release music, it can be important to paint a clear picture of who one is as an artist. When McKenzie has an established sound and has created an expectation for his audience, it can be interesting to release a duet. Memories Strife still has the Torfinn McKenzie sound of close vocals and well-written guitar parts. Looking forward, a second voice might feel refreshing to the listener.
Will the AI trigger inspiration or will it be distracting?
In the experience of the three participants, none agreed to the artificial tools being perceived as distracting. Artists like Ida Sofie Laland expressed a worry about the AI making the writing process too productive, to the point where she would feel that she was not in control. She experienced using the CLG as a kickstarter for writing lyrics. However, she did not experience the session to be rushed when using the tools. Laland believes this is because the AIs were solely used for inspiration.
Torfinn McKenzie had different opinions regarding AI tools. The CLG was perceived as good at creating ideas and associations, but where Scaler2 was not developed enough for him. This can be rooted in the fact that the genres he associates with were not an option within the plugin and that the creation of arrangement with his guitar holds such a big part of his music. For Laland and Raine, this was likely different as they first of all are singers, whilst McKenzie is a first a guitarist. The arrangement is personal to McKenzie, in the same way a melody can feel for a singer. I find it important to add that if there is no problem or desire, one would not have to use a tool. For McKenzie, the arranging of guitars and arrangement is a passion, and there might not be a need for tools to help the process at this point.
In Chapter 1.3 about limitations, I mention how writing with others can be challenging. One’s creativity can be affected by various factors, such as stress, sleep quality, comfortability and chemistry. Comparing the different sessions, also I was affected in multiple ways. My feeling of comfort increased with each interview. As the questions were fairly similar in the interviews, I felt more confident after doing it once or twice prior. In terms of chemistry, I approached the sessions differently. The artists were a mixture of people I knew and did not. It was therefore important for me to spend some time talking with the artists before the interviews. This can help to lower stress levels and make me more comfortable. In terms of creativity, I view the sessions similarly, which is often not the case. The sessions felt efficient and we did not get stuck, where we could not figure a part out. The closest call to feeling stuck was in the session with Raine, where we rewrote the chorus around 5 times. However, rather than feeling stuck, it felt more like having such a large amount of ideas, that one struggles to put it into structure. However, it eventually worked itself out and we chose one of the options.
Comfortability in a writing session is not always as easy as it sounds, as lyrics often are personal. "To have someone in the room [or on Teams] affects me. I get embarrassed", Laland explains. "I get embarrassed when trying to find the right words, explain situations or share something personal. It is something I need to work on". This can be viewed as a stress factor and can affect Lalands creativity during the session. We chose not to go into too much detail on the story behind the theme of the song because of this. I believe this is a valid point and understand the struggle of embarrassment. However, if can make it challenging for the cowriter, especially as she brought a theme for the song. The lyrics can end up contradicting without the transparency of what the song actually is about. I believe this issue was solved when Laland was to record vocals (alone), where she made changes to the lyrics to make it more cohesive.
Comfortability is important for the sessions and one should be prepared to get comfortable (Cole, 2018). A good example of this was my session with Raine. I did not know the artist before our session. However, we were both able to put this aside and write lyrics that we both feel emotionally connected to. The session was driven by a passion for the theme, as well as using references such as Olivia Rodrigo and Billie Eilish, which we both are fascinated by.
7 Conclusion
To conclude this thesis regarding songwriting and AI, the research question will be answered. This question is the focal point of this thesis. Through six interviews, three songwriting sessions, mixing, transcribing, researching, and writing, this question has been the main motivation during this whole process.
How do songwriters experience the use of AI in their songwriting process?
The artists participating in this thesis had a collective opinion regarding the CLG - The AI managed to trigger inspiration. By starting the writing session by presenting the CLG and generating text, inspiration was triggered. Ida Sofie Laland had a subject in mind for the session, and the word we used for the CLG was cold. This ended up also being the title of the song. One of the sentences produced by the CLG was "They say memories keep you company", which became the starting line of the chorus. Raine St. Frank did not have a subject in mind, but a feeling. We used the CLG to generate text from words like stop and angry. One of the lines generated by the CLG was, The next time I'll see you I'll gift you a mirror. The sentence became half of the pre-chorus, and triggered further lyrics, such as the continuing line: and maybe then you’ll see everything clearer. Torfinn McKenzie did not necessarily have a theme or an emotion in mind when coming to the session. We ended up using the search word sad, and the lyrics of Memories Strife came to life. Through only a few searches on the CLG in each session, followed by collecting sentences we liked, we were able to write three songs. The songs were written in two to four hours. Even though the CLG is fed only a fraction of what OpenAI's ChatGPT is, it serves its purpose. The limitation might even entail more relevant ideas.
Scaler2, used for generating chord progressions for these sessions, was met with different opinions. Raine St. Frank experienced it as a good starting point for a session. This was also my experience with using the plugin the first time. Even if one chooses not to use the chord progression or alter it during the process of writing, it functions as a good starting point. Laland somewhat agreed. However, after Cold was produced as a demo, she says she would want to play around with the chords more. Torfinn McKenzie, as mentioned before, did not find Scaler2 to be convincing enough. As a guitarist, he wants to create the chord progressions himself, even though it worked for Memories Strife.
Two of the sessions and five of the interviews were held over Teams. The interviews would ideally be held in person, as it is easier to develop trust and openness (Jacobsen, 2022, p. 165) that way. This was however only possible with one interview and one session, and the others were held over Teams. This brings me to the weakness of holding interviews in person. Even though it is ideal in terms of a conversational flow, it can be expensive in terms of travel costs (Jacobsen, 2022, p. 165). A good second solution was therefore to hold most of the interviews and sessions over Teams, with the cameras on. Being able to see each other makes it easier to read body language. This can be important, as one wants the participants to feel comfortable at all times. This could be hard to recognize if the interviews were held without the video option.
Something to consider if one wants to continue this research would be to double the number of writing sessions per participant. In Chapter 1 regarding limitation, I acknowledge that the need for creativity, and all that can affect one’s creativity, can affect how one views the session, such as the chemistry of the collaborators, sleep quality, and stress levels. However, the opposite of this can also be a hindrance to the artist's view of writing with artificial intelligence. If the writing session is filled with creativity, and good chemistry, and the session is considered successful, this can also have an impact on the artist's view on using AI in songwriting. The fact that the session in itself felt good and the musical result of the session was considered satisfying, could paint a picture of giving the credit to the AI, and not us. As I have not written songs with any of the artists before this project, I am also a new asset to the writing process.
Ideally one would therefore have one session with each artist without AI, and then follow with a second session where we write a new song, with AI. This would make the artists' opinions clearer. It would also hinder them in confusing me and the AI to be the reason for the session going well or not. Then again, this would be extremely time-consuming and one will never have two sessions that are the same, where the AI would be the only difference, and the creativity can still be affected by human factors. For further research, it would be interesting to see if this changed the participants' opinions of the use of AI in songwriting.
Another option for further developing this research could be by altering the CLG. As for now, the CLG can only be used by myself. For further research, it would be interesting to train the lyric generator on the participants' lyrics. The CLG would then write lyrics that possibly are closer to their way of writing.
The goal of this thesis was to continue the research on how artificial intelligence can be used as a tool. Through my background of writing bedroom/indie-pop with Rachel Meddings, the CLG, and Scaler2, I was inspired to take this process a step further. Through the work of this thesis and music creation relating to it, it was made possible to include three more artists in this way of writing. Through working with the three artists, we got to explore writing music where artificial intelligence has been involved in the lyric-making and the arrangement of chord structures. Through six qualitative interviews, the artists shared their prior experiences and expectations before the writing session and the experience they had during the writing sessions afterward. As well as this, the artists shared their thoughts regarding ownership and how they see the AI songs fit their catalogue.
Bibliography
APA Dictionary of Psychology: qualitative research. (05/19/2018). In American Psychological Association. Online <25/01/2025> from https://dictionary.apa.org/qualitative-research
Artificial Intelligence (U.Å) In the Cambridge Dictionary. (20th of March 2024). Cambridge University Press.
https://dictionary.cambridge.org/dictionary/english/artificial-intelligence
Bennett, J. (2011). Collaborative Songwriting: The Ontology of Negotiated Creativity in Popular Music Studio Practice. Journal on the Art of Record Production, (Issue 05),1-18. https://www.arpjournal.com/asarpwp/collaborative-songwriting-%e2%80%93-the-ontology-of-negotiated-creativity-in-popular-music-studio-practice/
Bjørkeng, P. K. (2018). Kunstig Intelligens: Den Usynlige Revolusjonen. Vega Forlag
Bloomfield, T. (1993). Resisting Songs: Negative Dialectics in Pop. Popular Music, 12(1), 13–31. http://www.jstor.org/stable/931256
Blume, J. (2022). What Makes the Perfect Demo. Music World. https://www.bmi.com/news/entry/what-makes-the-perfect-demo
Botpress Community (July 11th, 2023). Are There Any Legal or Copyright Concerns When Using ChatGPT-Generated Content? https://botpress.com/blog/are-there-any-legal-or-copyright-concerns-when-using-chatgpt-generated-content
Cave, N. (2023). ChatGPT Making Things Faster And Easier. The Red Hand Files. https://www.theredhandfiles.com/chatgpt-making-things-faster-and-easier/
Chomsky, N. (2023). Noam Chomsky: The False Promise of ChatGPT. The New York Times. https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chatgpt-ai.html
Clancy, M. (Eds.) (2023). Artificial Intelligence and Music Ecosystem (1st Edition). Taylor and Francis
Cole, C. (2018). Navigating the Co-writing World: The Do’s and Don’ts of a Grat Music Relationship. Cari Cole. https://caricole.com/navigating-the-cowriting-world-the-dos-and-donts-of-a-great-music-relationship/
Cook, Nicholas (2018). Music as Creative Practice. Oxford University Press
DiMatteo, L.A. & Poncibó, C. & Cannarsa, M. (Eds.) (2022). The Cambridge Handbook of Artificial Intelligence: Global Perspectives on Law and Ethics. Cambridge: Cambridge University Press. https://doi.org/10.1007/978-3-030-72116-9
Edmondson, A. C. (2019). The Fearless Organization: Creating Psychological Safety in the Workplace for Learning, Innovation, and Growth. Hoboken, NJ: John Wiley and Sons
Experience (U.Å) In the Cambridge Dictionary. (26th of March 2024). Cambridge University Press. https://dictionary.cambridge.org/dictionary/english/experience
Freshbots (2023). AI Song Lyrics Generator. Freshbots. https://www.freshbots.org/
Frith, S. & Goodwin, A. (Eds.) (1990). On Record. Routledge
Galletta A. (2012). Mastering the Semi-Structured Interview and Beyond: From Research Design to Analysis and Publication. New York University Press, New York.
Gjøsund, P. & Huseby, R. (2020). To eller flere (4th edition). Cappelen Damm AS
Harhaug, L. (2020). Goodbye Intuition. Norges Musikkhøyskole: Norwegian Academy of Music. Online: https://www.researchcatalogue.net/view/974962/979502
Henriksen, D., Richardson, C., & Shack, K. (2020). Mindfulness and creativity: Implications for thinking and learning. Thinking skills and creativity, 37, 100689. https://doi.org/10.1016/j.tsc.2020.100689
Hetler, A. (2023). What is Generative AI? Everything You Need to Know. TechTarget. https://www.techtarget.com/whatis/definition/ChatGPT
Hope, A. (2023). How To: What Is A Pop Song Structure? Music Gateway. https://www.musicgateway.com/blog/how-to/what-is-a-pop-song-structure
Houghton, B. (2023). AI and how it will affect musicians, songwriters, authors, and creators.
Hypebot. (3rd of October 2023) Retrieved from https://www.hypebot.com/hypebot/2023/03/ai-and-how-it- will-affect-musicians-songwriters-authors-and-creators.html
Iveland, K. L. (2024). Lyrics, Voices and the Stories They Tell [Doctoral Dissertation]. Universitetet i Agder
Jacobsen, D.I. (2010). Forståelse, beskrivelse og forklaring (2nd edition). Høyskoleforlaget
Jacobsen, D.I. (2022). Hvordan gjennomføre undersøkelser (4nd edition). Cappelen Damm Akademisk
Kaufmann, G. (2015). Hva er: Kreativitet. Universitetsopplaget
Killgore, W.D.S. (2010). Effects of Sleep Deprivation on Cognition. Science
Direct. https://doi.org/10.1016/B978-0-444-53702-7.00007-5
Krishnarai, N., Bae, J., Meng, X., Heard, Y. & Cardwell, C. (2023). Benefits of AI: Increase Efficiency, Enhance Safety, and Improve Reliability. Credera. https://www.credera.com/en-us/insights/benefits-of-ai-increase-efficiency-enhance-safety-and-improve-reliability/
Lyng, C. (2021). Awful Crime [Song]. Awful Crime. Darling Records
Marr, B. (2023). A Short History of ChatGPT: How We Got to Where We Are Today. Forbes. https://www.forbes.com/sites/bernardmarr/2023/05/19/a-short-history-of-chatgpt-how-we-got-to-where-we-are-today/?sh=61bd7ce2674f#open-web-0
McKinsey & Company. (2024). What is Artificial General Intelligence (AGI)? McKinsey & Company. https://www.mckinsey.com/featured-insights/mckinsey-explainers/what-is-artificial-general-intelligence-agi#/
Meddings, R. (2022). The AI Songwriter: Coding a Co-Writer [Master’s Thesis]. BIMM Birmingham
Miranda, M. C., Barbosa, T., Muszkat, M., Rodrigues, C. C., Sinnes, E. G., Coelho, L. F. S., Rizzuti, S., Palma, S. M. M., & Bueno, O. F. A. (2012). Performance patterns in Conners' CPT among children with attention deficit hyperactivity disorder and dyslexia. Arquivos de Neuro-Psiquiatria, 70(2), 91–96. https://doi.org/10.1590/S0004-282X2012000200004
Nettskjema (February 14th, 2024). Nettskjema. Retrieved from February 14th, 2024.
https://nettskjema.no/?redirectTo=/user
Nielsen, F.V. (1998). Almen Musikdidaktik (2nd edition). Fredensborg: Christian Ejler’s forlag
OpenAI. (2023). ChatGPT (March 28th version) [Large Language Model]. https://chat.openai.com/
Oxford Learners Dictionaries (28th of February 2024). Methodology. Oxford Learners Dictionaries. https://www.oxfordlearnersdictionaries.com/definition/english/methodology?q=methodology
Plugin Boutique. (April 26th, 2023). Scaler 2. Plugin Boutique. https://www.pluginboutique.com/product/3-Studio-Tools/93-Music-Theory-Tools/6439- Scaler-2
Priest, M. (2024). Large Language Model Explained. Boost.AI. https://boost.ai/blog/llms-largelanguagemodels/#:~:text=Pre%2Dtraining%3A%20The%20model%20is,next%20word%20in%20a%20sentence
Use (U.Å) In the Cambridge Dictionary. (24th of March 2024). Cambridge University Press https://dictionary.cambridge.org/dictionary/english/use?q=use+
Warner, T. (2018). Pop Music: Technology and Creativity - Trevor Horn and the Digital Revolution. United Kingdom: Taylor & Francis.
Weatherbred, J. (2024). How to Keep Your Art Out Of AI. The Verge. https://www.theverge.com/24063327/ai-art-protect-images-copyright-generators
Wiener, A. (2023). Holly Herndon’s Infinite Art. The New Yorker. https://www.newyorker.com/magazine/2023/11/20/holly-herndons-infinite-art
Appendix B Guide for the first interview
Note that the interviews were semi-structured. The interviews were not limited to the questions below.
First interview
-
Focusing on your prior releases, how do you describe your writing and musical sound?
-
Have you written songs with AI before?
-
If not, what are your expectations?
-
-
What are your thoughts on using AI for writing songs?
-
Do think music made with AI will lack some human aspect?
-
What are your expectations for this session?
Appendix C Guide for the second interview
Note that the interviews were semi-structured. The interviews were not limited to the questions below.
Final interview
-
Short introduction of yourself, name, from, musical background.
-
How would you describe the sound of your previously released music?
-
How was your experience using the lyric generator?
-
If their mind has changed, why?
-
-
How was your experience using Scaler2?
-
What did you like about using CLG or Scaler2?
-
What did you not like about using CLG or Scaler2?
-
Do you think the song we created fits with your already released music?
-
Did the AI shorten the process of writing songs? More productive or distracting?
-
Do you feel “the song” lack some human aspect/authenticity?
-
Do you feel the same amount of ownership to “the song” as your other music?
-
Why, why not?
-
-
Will you be using AI when writing songs in the future?
Appendix C Guide for the second interview
Note that the interviews were semi-structured. The interviews were not limited to the questions below.
Final interview
-
Short introduction of yourself, name, from, musical background.
-
How would you describe the sound of your previously released music?
-
How was your experience using the lyric generator?
-
If their mind has changed, why?
-
-
How was your experience using Scaler2?
-
What did you like about using CLG or Scaler2?
-
What did you not like about using CLG or Scaler2?
-
Do you think the song we created fits with your already released music?
-
Did the AI shorten the process of writing songs? More productive or distracting?
-
Do you feel “the song” lack some human aspect/authenticity?
-
Do you feel the same amount of ownership to “the song” as your other music?
-
Why, why not?
-
-
Will you be using AI when writing songs in the future?
Appendix D Lyrics for Cold
© Celine Lyng and Ida Sofie Laland 2023
Verse
Feeling sad when snow is falling
Outside my window
I know the world is always changing
But our memories remain
Verse
I can't help but feeling kinda silly
Maybe I am immature
I could stand outside your wooden door
but can I make it in the cold
Chorus
Some say that memories keep you company
but could they possibly comfort me
The fog is stubborn
I'm stuck in between
you’re changing me
you’re changing me
Verse
When the morning comes but there’s no light
Oh I think of you cause you're my lantern
Give me a sign so I can rest assured
that I am welcomed to your home
Chorus
Some say that memories keep you company
but could they possibly comfort me
The fog is stubborn
I'm stuck in between
you’re changing me
you’re changing me
Bridge x3
Maybe we could sit together
Put on our fuzzy sweaters
Echoes of the past are ringing
I feel warm from reminiscing
Chorus
Some say that memories keep you company
but could they possibly comfort me
The fog is stubborn
I'm stuck in between
you’re changing me
you’re changing me
Appendix E Lyrics for The Things I Know
© Celine Lyng and Raine St. Frank 2023
Verse
You tell me to save more, but I barely spend
You tell me to work out and get better health
you never say sorry when making me cry
but fuck you´re quick at handing out advice
Pre-chorus
The next time I see you I’ll gift you a mirror
and maybe then you’ll see everything clearer
Chorus
Stop telling me these things that I know (bla bla bla bla)
Stop telling me these things that I know (bla bla bla bla, bla bla bla bla)
You keep telling me to handle my emotions
I reckon I just need to be more cautious
Stop telling me these things that I know
Stop telling me these things that I know
Verse
You say life's just like that, goes up and down
Say it'll be better the next time around
You ask me “How are you”, sounds nice and sweet
I keep oversharing, things you´ll later use against me
Pre-chorus
The next time I see you I’ll gift you a mirror
And maybe then you will see everything clearer
Chorus
Stop telling me these things that I know (bla bla bla bla)
Stop telling me these things that I know (bla bla bla bla, bla bla bla bla)
You keep telling me to handle my emotions
I reckon I just need to be more cautious
Stop telling me these things that I know
Stop telling me these things that I know
Bridge
(Improvisation)
Chorus
Stop telling me these things that I know (bla bla bla bla)
Stop telling me these things that I know (bla bla bla bla, bla bla bla bla)
You keep telling me to handle my emotions
I reckon I just need to be more cautious
Stop telling me these things that I know
Stop telling me these things that I know
Appendix F Lyrics for Memories Strife
© Celine Lyng and Torfinn McKenzie 2024
A
Keep regrowing the garden but the leaves always wither
Buy the same flowers that die early in winter
I sing the same songs but now the lyrics feel bitter
I know you’re not here but still, your voice lingers
B
The wind carries the burden of an empty life
ripples the walls, holding on, where memories strife
there’s always a table, set for two
What’s the point if I’m not here with you
C
(Instrumental)
A
Do you remember our first summer together
I gave you flowers, you said you'd keep them forever
We danced all night to our favorite song
With no care in the world, 'cause with you, my love what could go wrong
B
The wind carries the burden of an empty life
Ripples the walls, holding on, where memories strife
There’s always a table, set for two
What’s the point if I’m not here with you
C
(Instrumental)