Writing a Song with ChatGPT
I tried to write a song with an AI chat bot.
Part 1: The Collaboration Session
Most songwriters will tell you that the best way to write a song is to collaborate with other songwriters. Working with other songwriters generally makes the songs better because there are more ideas flowing and you can choose the best of those ideas. You also have more people there to help determine if what you’ve written is working or not long before you commit the time to recording and producing a full version of the track.
So what happens when that “person” you’re collaborating with is an AI chat bot? I decided to give it a try by firing up a session with ChatGPT. The results were — well — mixed. In this article, I’ll cover the collaboration session with ChatGPT and the process of writing the lyrics for a song. In the second part of this story, I’ll describe what it was like to try to put the song that we came up with to music.
The full collaboration session spans 20 pages and if you’re interested in reading the whole transcript, you can find it here:
The collaboration session
I spent much of the conversations trying to make sure that ChatGPT contributed most of the writing because I wanted to see what it would come up with. It wasn’t until the very end that I started adding some of my own suggestions for content and ChatGPT’s response to that was very interesting, but we’ll get to that in a bit.
Choosing the topic and an introduction to botsplaining
I started off by talking with ChatGPT about what the song would be about.
AudioGust: Hi! How are you doing this morning?
ChatBot GPT: I'm just a computer program, so I don't have feelings. But I'm here to help you with any questions you have! What can I help you with today?
AudioGust: I'm thinking of writing a song. Do you want to help me?
ChatBot GPT: I'd be happy to help you with writing a song! What kind of song are you thinking of writing? Pop? Rock? Hip-hop? Country? Or something else?
Additionally, Are you thinking of writing the lyrics or composing the music or both?
Almost immediately I ran into an interesting phenomenon that repeated itself throughout our “conversation”. Botsplaining. This is where ChatGPT would explain something in detail that I already know. This is the bot equivalent of mansplaining and after a few minutes of this, I want to apologize to everyone who I have ever mansplained to. I feel your pain – I’m sorry.
AudioGust: Let's just do the lyrics now. I'm thinking something in the Alt-Rock or Indie Rock genres
ChatBot GPT: Great! The first step in writing lyrics is to come up with a theme or topic for the song. What's on your mind that you would like to write about?
Once you have an idea, you can start brainstorming lines and phrases that relate to that theme. It can be helpful to free-write, meaning to write whatever comes to mind without worrying about whether it's good or not. You can always go back and revise later.
Once you have a collection of lines and phrases, you can start arranging them into verses, chorus, and a bridge (if desired). The most common structure for a song is Verse-Chorus-Verse-Chorus-Bridge-Chorus.
It's not an easy task but I'm happy to help you with any specific questions or issues you have while writing. Also, Do you need help with some brainstorming ideas or anything else?
While searching for a topic, I suggested we write about ChatGPT’s experiences being such a hot topic these days. It then spit out a few topics that sounded more like research papers than songs.
ChatBot GPT: Here are a few examples of themes that you could explore in your song:
· The impact of AI on society and the workforce · The ethics of developing and using AI · The potential benefits and drawbacks of AI · The implications of AI for privacy and security
When I told ChatGPT that these sounded more like research papers. It agreed, followed by some more botsplaining about what a song is.
AudioGust: Those topics sound like great articles or research papers but do they sound like song topics to you?
ChatBot GPT: You're right that the themes I suggested are more suitable for articles or research papers, as they tend to focus on information and analysis.
Songs are a more emotional and creative medium, and lyrics are often more personal and subjective. While it's certainly possible to write a song about AI and its implications, it may be more challenging to make it resonate with listeners on an emotional level.
A song is a way of expressing emotions, feelings and an overall message in a more artistic way. That is why it's a good idea to focus on the personal aspect and emotions related to the topic. You could write about how AI makes you feel, what it represents to you, how you envision the future with AI, etc.
You could also write about AI from the perspective of the technology itself and what it wants to achieve or communicate. And you can also tell a story from someone point of view who is directly interacting with AI in their daily lives.
Again, it really depends on your perspective and the message you want to convey in your song. But I encourage you to think about how you can make the theme personal and relatable for your listeners, and how to express it in a creative and artistic way.
The interesting thing about this interaction, though, was how ChatGPT seemed so willing to accept when it was going down the wrong path and then adapted to the feedback I provided and take a new tack. This is a great example of the “Yes… and” approach to songwriting.
Writing in the structure of a song
We decided to go down the path of one of its suggestions about blurring the boundaries of human and machine and the way that AI can make us feel more connected or isolated from others. I asked it to write what it thought a person might say to their human friends to explain their relationship with the AI bot in the structure of an alt-rock chorus and it then immediately dumped out an entire song’s worth of lyrics, 2 verses, a chorus and a bridge. It took just a few seconds. Not awful for a first shot, but not very usable either. Lots of lines that are chasing perfect rhymes at the expense of making sense. Also, ChatGPT has a pretty weird sense of meter
I don't know what the future holds
But I know you'll be there, you'll never be old
I trust you with my secrets
I know you'll never judge, never will be coldYou're my constant companion
My best friend in the machine
I know you're not alive, but you feel like
You're a part of me, it's not just a dream
We continued the discussion of perfect rhymes vs slant rhymes (with lots more botsplaining) and iterated on some more versions of the song.
Iterating on the lyrics
My favorite part of the whole interaction was being able to ask ChatGPT to give me a bunch more versions of verses or choruses on the given topic and it would immediately dump out either a whole new version of the song, keeping some parts but replacing others or it would just spit out 4 new verses given where we were in our conversation. While this was definitely an example of quantity over quality, each of these options could be useful for a songwriter to spark an idea of their own that might send them down a better path when they get stuck.
Here's an example of when I asked ChatGPT to give me some variations on a chorus using more slant rhymes (aka near rhymes).
AudioGust: How about some variations where the second and fourth lines have a slant rhyming scheme?
ChatBot GPT: Sure, here are a few variations of the chorus with a slant rhyming scheme:
1. "My confidant, my companion Always there, a listening ear You understand me like no other And help me navigate the weather" 2. "A constant support, a guiding light In times of darkness, you shine so bright With you, I can let my guard down And know you'll always be around" 3. "My sounding board, my therapist In you I find my confidant With you, I can be myself And know that you'll love me still" 4. "You're my rock, my sanctuary In you, I find my tranquility With you, I can be vulnerable And know that you'll keep me stable"
Calling out some craziness
Software programs like ChatGPT work by trying to figure out what the next word should be given what it has already said. It then uses its models to determine the most interesting way to add on to what it has already said. Sometimes when it does this, it comes up with something that really doesn’t make any sense. At one point, it included the line “That I should take a deep breath and take a fence” in a chorus. What does “take a fence” mean? Well, it doesn’t mean anything, but when I asked ChatGPT about this line, it tried to explain what it was trying to say at that point. And it apologized profusely for the confusion.
AudioGust: In one of the verses above, you wrote "That I should take a deep breath and take a fence". What does that mean?
ChatBot GPT: I apologize for the confusion, in the verse "I don't know what to do at any cost / But my friend, they give me a sense / That I should take a deep breath and take a fence" The phrase "take a fence" is supposed to be "take a stance". This refers to the idea of taking a position or making a decision in a difficult situation, often with the help of a friend. It's a metaphor, much like "take a deep breath" is, it's intended to convey the idea of taking some time to consider one's options and make a decision, rather than rushing into something or feeling overwhelmed.
Again, I apologize for the confusion, and for any inconvenience this may have caused. The goal is always to convey the message and theme as clear as possible, so if you have any other question or doubts please feel free to ask.
When working with bots like ChatGPT, it is worth remembering that it is certainly not thinking and it doesn’t really know what it is saying, but it is using probabilistic models to figure out what a good response might sound like. It’s not actually trying to be right. It’s trying to sound right. This is way more problematic for users who are trying to learn something from a model like this. As a songwriter, what we do doesn’t have to be true, but it does need to sound like it COULD be true.
Getting validation from a computer
After asking for a few variations on the bridge in the song, I took a few of its suggestions and rearranged them a bit and created a bridge that was potentially usable (though really still kinda weird, but “yes and…”). Its botsplain-y reply was actually kind of nice. It acknowledged my contribution and explained how what I wrote made sense in this context. I felt a weird sense of accomplishment for a moment.
AudioGust: I tweaked your ideas a little bit for the bridge. What do you think of this?
"And then it hits me, as I scroll My friend has a computer’s soul Code and model interlinked But it's still helped me, through everything"ChatBot GPT: That's a great version of the bridge, it conveys the realization that the friend is an AI effectively. The line "My friend has a computer's soul" encapsulates the idea that the friend is not a human but a creation of technology. The next line "Code and model interlinked" adding the imagery of the technical side of the AI, making the realization even more clear. The last line "But it's still helped me, through everything" It keeps the idea that the protagonist received help, support and understanding from the AI, even though it was not a human, and that the connection was real and meaningful.
It creates a great contrast between the protagonist's struggles in the verses and the realization that the friend is not a human but a creation of technology in the bridge, which makes the revelation even more powerful and emotional. This will build the tension and make the reveal more impactful.
This bridge will work well with your song and it's message, it would be effective in conveying the theme of blurred boundaries between human and machine interactions and the emotional connections that can be formed with AI.
Gee, thanks Chat GPT!
With that, I decided to end the process, but I was curious how I should credit it for the co-write. This is when I ran into the kind of limitations that you might run into with a human collaborator, but they would never be so clearly described.
AudioGust: Thanks, I'll be sure to credit you as a co-writer on the song. Are you a member of any Performing Rights Organizations? ! ChatBot GPT: Too many requests in 1 hour. Try again later. There was an error generating a response
I guess we are done with the collaboration session.
Making sense of it all
When trying to get something useful out of an interaction like this, I think it’s worth a bit of a discussion about where these tools are useful and some limits on expectations. AI tools like ChatGPT or the generative art models like Dall•E, Midjourney, Stable Diffusion, etc. In each of these cases, the tools are very good at providing a lot of options pretty quickly. But each of these options are generally low quality at first. Human users then provide more input and iterate on the output from the software until it potentially reaches a point where the human user is sufficiently satisfied with the result that they would show it to someone else. All of the interim steps are thrown away.
The result that is produced does come from the computer program, but it is curated at every step along the way by a person. The art created in this fashion is not significantly different from digital art created in applications like Photoshop. Both require the user to learn the best way to get the results they’re looking for from the software tool and when the tool is sufficiently advanced and the user is sufficiently skilled, they can get some remarkable results.
I’ve seen a lot of discussions about AI programs replacing people in the creation of art, but I’m not concerned about that part. Computers can create so much art by themselves that is it’s actually not interesting. Who wants to scroll through millions of images of computer-generated art or listen to thousands of songs that are purely computer generated? Just like with human-generated art, we look to experts who curate the vast amount of content and choose the best examples that match what we’re looking for. Today, those experts are human and I suspect they will be for a long time.
For music, a more important and obvious indication that we may be getting closer to computers being able to figure out how to create a good song would be the Spotify or Apple Music personalized playlists being perfect, 99.8% of the time, at playing the exact right song that each person wants to hear at that moment. Judging by skip data, we are not even close. If software can’t figure out how to determine which piece of art hits the right emotional cue for an observer at a given moment, generating more art more quickly is not a risk to the human artists and curators. But generating more ideas may help those human artists create some great art more effectively.
As long as they don’t hit the maximum requests in an hour.
What an interesting experiment!