Will ≥1% of Spotify's streams in 2024 consist of AI-generated music?
59
272
1k
2025
14%
chance
Suno v3 launched
Udio launch date

Some people are claiming we're facing "a modern Napster moment" in music, from AI-generated songs of various sorts. Will the hype continue, is it exaggerated, or will perhaps the industry shut it down?

Resolves YES if credible news sources report this being the case, or numbers/estimates that heavily indicate that this is the case. Note that this leaves the burden of proof on the YES side.

I will not bet in this market. So I am leaving other resolution criteria open to my best judgment, in case needed.

I expect a YES to be able to resolve early, but a NO to happen at the end of Q1 2025 if no signs of YES show up.

Get Ṁ600 play money
Sort by:

I realize sharing these is the modern version of "listen to my mixtape," but it's just so fun...
Streamin' On Repeat

(this is the first result Suno v3 gave when prompted with the title of this market)

reposted

This question has gotten a lot more interesting in the past week 👀

sold Ṁ86 NO

Holy crap Suno has gotten really good

Considering the clarifications added, I say NO.

(*) even though Spotify stated that they are not banning AI songs on its platforms (1), the CEO noted that it is against AI music that uses other artists' voices, such as Drake and the Weekend example.

(*) Those songs happened to be the ones that so far became popular in streaming. If we search for top-generated songs, these are the ones that contain voices from other artists, like Heart on My Sleeve, which had 600,000 streams on Spotify (2).

(*) On the contrary, according to experts, the rest of the other types of AI songs, which are 100 million among popular platforms such as Spotify (3), are rarely played because they end up mixing with other million types of music.

(*) This year, Spotify deleted tens of thousands of AI for suspicious streaming activity, aka "fake listening" (4), and experts say the increase in the amount of AI music will lead to major issues in fake listening, making it harder to know if these types of songs are popular or not.

Source

(1) https://www.bbc.com/news/technology-66882414

(2) https://www.greataiprompts.com/guide/best-ai-songs/

(3) https://www.wired.com/story/ai-generated-music-streaming-services-copyright/

(4) https://www.wired.com/story/spotify-ai-music-robot-listeners/

predicts NO

Big ol' NO from Gwern.

> (1) the copyright situation is much scarier. Music comes with all sorts of bizarre complex IP laws and rights that don't apply to text or images, and this is because of (and empowers) groups like the RIAA. Every commercial startup doing music-related generative work seems to keep things very quiet, as blackbox as possible, restricting to licensed datasets when possible, and definitely not releasing stuff. (2) lack of major hobbyist interest - people may like to listen to music, but a lot of it is parasocial, there's superstar winner-take-all effects, and people don't generally want to make their own music the way they do their own images & text - think about porn/hentai as a huge driver of image synthesis. There is no porn/hentai of music generative models. (3) relatively demanding compute compared to other modalities. Audio is just very, very bulky: a piece of music is made out of millions of individual datapoints (X frequencies times Y milliseconds), while you can make a recognizable image with just 64x64=4096 pixels. Jukebox wasn't as demanding as GPT-3-175b was... but it's still a lot more compute than most people want to spend.

An important clarification: does the music need to be entirely AI-generated?

I could imagine a world where many beats/samples/autotune is AI-generated but humans still perform it. Or AI-generated backing tracks with a human singing.

I think of this market as entirely AI-generated and I'm curious for @HenriThunberg 's take before I predict.

(I also assume you mean full year 2024, vs. this time next year)

@OllieBase thanks, I wanted to add some of these important clarifications. Will try to build from the linked tweet and flesh out from there.

I think the line will be too blurred if we include beats/samples/autotune. To me, there are two main cases I think are interesting:
1) AI-generated lead vocals, with or without human assistance. While not entirely AI-generated, I think it best represents the Drake song in the description tweet.
2) Instrumental music that is completely AI-generated. I could definitely see this happening, and being an important and relevant development, for playlists like "Lo-fi beats"/"Piano music to study to"

That means I would want to exclude
3) Human lead vocals + AI-generated backing/samples/...
4) Lyrics written by AIs, but performed by humans
5) Music written by AIs, but performed by humans

Happy to add more cases on either side if convinced it's useful.

Regarding timelines, yes I assume figures for all of 2024 to be easiest to lay our hands on rather than a snapshot picture of April 2024.