Thread by @ednewtonrex

I analyzed a bunch of Suno's output, and found it generates music that closely resembles well-known, copyrighted music: melodies, chord progressions, lyrics, instrumental parts, and styles. This raises serious questions about what it’s trained on.

Here are some examples 👇🧵

(Link to my full piece in @musicbizworld further down)

I've been able to generate music in the style of most artists I've tried. Here's Eminem. Suno has some prompt filtering, but, like other gen AI systems, these are easy to get around with mis-spellings etc.

youtube.com/shorts/nkN9XYF…

More telling than stylistic similarity, though, is melody. Here's what I got when I entered the lyrics for the 'Galileo' bit from Bohemian Rhapsody.

Here's the start of what Suno generated when I provided the lyrics to Tracks of My Tears by Smokey Robinson & The Miracles, compared to the original. The first five pitches are the same.

The chances of this happening by chance are... low.

(I transposed all written examples to C major for ease of comparison.)

Generating music to the lyrics of Oasis' Wonderwall, I got a couple of very similar patterns to the original.

Same for Don't Look Back in Anger (I'm an Oasis fan).

A couple of differences here (marked), but clear similarity.

Generated chord sequences resemble copyrighted music, too. When trying to generate 80s dance-pop, I got this track, whose chords are very similar to Gloria Gaynor’s I Will Survive.

Original: Cm - Fm - Bb - Eb - Ab - Dm half dim. - Gsus - G
Suno: Cm - Fm - Bb - Eb - Ab - Fm - G - G

I also found that, when using Ed Sheeran's Castle on the Hill lyrics, I got very similar chords to the original.

Original: C - C - F - F - Am - Am - G - G
Suno: C - C - F - F - C - C - G - G

(C for Am is pretty much the simplest chord substitution you could have here.)

That chord sequence also featured elsewhere in Suno's output, with a bass line similar to Castle on the Hill, too (they share the characteristic third scale degree approach to chord IV, and the repeated notes on each chord).

And the vocalist sounds familiar...

And when using the Dancing Queen lyrics, I got a characteristic chord sequence from Abba's original.

IMO, either this is all chance, or Suno is trained on copyrighted music. I can't think of another alternative (though I'm open to suggestions here).

If it's trained on copyrighted music, Suno may have licenses in place to do so. But one of Suno's investors indicated they didn't have licenses when they invested, and Billboard's sources recently confirmed they don't have licenses in place with the three major labels. As far as I can tell, there are similarities here to all three of the major labels' music.

billboard.com/pro/ai-music-c…

So we should be asking what Suno train their models on. They don't reveal their training data, so the only way we can try to work it out is this kind of red-teaming.

There are many (myself included) who believe that training generative AI models on copyrighted work without a license is copyright infringement.

I hope Suno aren't training their models on copyrighted work without a license, and that they will clarify what their models are trained on. Licensing training data has to be the way forward for generative AI companies. Without licensing, these models "substantially dilute the royalty pools that are paid out to artists", as ~250 artists and songwriters said this weekend.

There are more examples in this piece I wrote for @musicbizworld: musicbusinessworldwide.com/suno-is-a-musi…