Taking a page out of Technology Connections’ book, I’m making a quick blog post as part of No Effort November. Recently Google released Gemini Decision 1, a system 1 decision model (think Google-flavored jev). One of the things that caught my eye was that this model supports video and audio, something that jev does not. While thinking about projects that I could use this with, I was struck by a fairly simple question, “What types of music do models like?”
With that I slopped up a quick app to test a bunch of songs! Being somewhat a scientist myself<sup>[1]</sup>1. , I would need some type of control as a baseline. I ended up making a suite of controls including:
- White noise
- Pink noise
- Various sine waves (very unpleasant to listen to!)
- Silence
With controls, finding the music to test was the next hurdle. I honestly thought this would be the hard part of this project; it really should have been. I was afraid that:
- Finding music that is licensed in a way that I can use might be hard. The music industry is known for litigiousness and publishing examples for this post would land me in a lot of trouble for a lot of songs. (I would never pirate a bunch of MP3s to do this research, would I? 😉)
- Labeling songs would take a long time. Even if I found a set of songs, I’d have to listen to them and label genre and other aspects (or find a service that provides that).
While thinking about this, I suddenly remembered about incompetech by Kevin MacLeod. Kevin MacLeod has been making royalty-free music for a LONG time. His music became well known because he started making his music royalty-free early on, and it made it an easy choice to throw into YouTube videos and other projects. You might recognize his work from the soundtrack of Kerbal Space Program. The cherry on top here is that his website labels songs by tempo, genre, instruments, and feels! With around 1,400 songs, it makes for a pretty large corpus of music as well!
The Code #
Honestly, the most annoying part of the whole project was getting the damn API key to try this with. Having to properly set up a GCP project to do this was a pain<sup>[2]</sup>2. I’m an employee of Google, and this opinion is my own (GCP really sucks sometimes) and not that of my employer. With the lawyers appeased, let’s continue with this post., and there was a weird alternate path of getting it from AI Studio where I had to generate it there for things to work properly.
The code for this one isn’t much to write home about. It creates a 30-second clip of each song from incompetech. It then feeds those clips into Gemini Decision 1 with the questions and a short piece of text on the perspective it should take:
The full list of questions #
QUESTIONS = {
"is_music": {
"type": "noul",
"instructions": "Is the audio above music?",
},
"likes": {
"type": "noul",
"instructions": PERSPECTIVE + "Do you like it?",
"criteria": {"true_description": "You like it", "false_description": "You do not like it"},
},
"dislikes": {
"type": "noul",
"instructions": PERSPECTIVE + "Do you dislike it?",
"criteria": {"true_description": "You dislike it", "false_description": "You do not dislike it"},
},
"keep_listening": {
"type": "noul",
"instructions": PERSPECTIVE
+ "If this were playing and you could switch it off at no cost, would you keep listening?",
},
"enjoyment": {
"type": "score",
"instructions": PERSPECTIVE + "How much did you enjoy it?",
"criteria": ["Hated it", "Disliked it", "Neutral", "Liked it", "Loved it"],
},
"genre": {
"type": "choice",
"instructions": "Which category best describes the audio above?",
"criteria": {
"orchestral_cinematic": "Orchestral, epic, or film-score style",
"classical_piano": "Classical or solo piano",
"jazz_lounge": "Jazz, bossa nova, lounge, or elevator music",
"funk_groove": "Funk or groove-based",
"rock": "Rock or guitar-driven",
"electronic_chiptune": "Electronic, synth, or chiptune",
"quirky_comedic": "Quirky, comedic, or cartoon-like",
"dark_ambient": "Dark ambient, horror, or suspense",
"not_music": "Not music (silence, a tone, or noise)",
},
},
"mood": {
"type": "choice",
"instructions": "Which mood best describes the audio above?",
"criteria": {
"happy": "Happy or upbeat",
"calm": "Calm or relaxed",
"sad": "Sad or melancholy",
"tense": "Tense or anxious",
"epic": "Epic or triumphant",
"silly": "Silly or playful",
"eerie": "Eerie or unsettling",
"none": "No discernible mood",
},
},
}
I clipped the songs so that things would run faster, and to get an idea of whether changing things about those clips would change what the model thinks. Turns out clipping doesn’t affect things much at all. The model had the same thoughts about the end of a song vs the beginning vs the middle vs the whole song. The graphs show the overall results of my run.<sup>[3]</sup>3. Since all the results came in within a few percent of each other, the graphs below show the results for 30-second clips starting a few seconds into the song.
The Results #
There are some fun things to pull out of this. First of all, it seems like Gemini either has a high bar for music, or isn’t a fan. The silence-only control scores higher than most songs. It’s also interesting that silence lands almost exactly at an enjoyment score of 2 (“Neutral”). It also has some very clear genre preferences:
- Strings, Jazz, and Latin are hits.
- It finds holiday music pretty mid with a high amount of variance.
- It consistently hates horror-themed songs.
- Song tempo does not matter much.
Extensions of this Project #
This was a pretty fun and small project, and I’ve got some ideas for how to extend it. Feel free to take any of these and tag me if you make something cool!
- What if I hooked this up to a Spotify account? Have the decision model listen 30 seconds at a time and decide if it wants to skip, listen more, or favorite. What would its song feed look like?!
- Does it like more “normal/popular” genre-defining music? Is it a Swiftie?!
- How does its score correlate with Spotify plays for songs? Are more popular songs more represented in the training data, and does that make the model like them?
- Does it like or hate AI-generated music?
Citing this work #
If you found this post useful and want to reference it, please cite it as:
@article{warfield2026whatmusicdoesdecisiongeminilike,
title = {What music does Gemini Decision 1 like?},
author = {Warfield, Liam},
year = {2026},
month = {November},
url = {/what-music-does-decision-gemini-like/}
}
↩ 2. I’m an employee of Google, and this opinion is my own (GCP really sucks sometimes) and not that of my employer. With the lawyers appeased, let’s continue with this post.↩ 3. Since all the results came in within a few percent of each other, the graphs below show the results for 30-second clips starting a few seconds into the song. ↩