I Replaced Myself with AI, Could You Tell?

Check out the audio podcast version of this post on Apple Podcasts or Spotify.

Do you think you can tell the difference between me playing a sax solo and AI? I set up an experiment using the AI music app SUNO, and the results are freaky. I wanted to see if I could get AI to play the saxophone like me, and this is the tool I used.

The other day I recorded a video of myself playing over a new backing track from BetterTrax for the tune St. Thomas. I played the melody and then improvised a solo.

SUNO has this feature where you can extend a song that you upload. So I uploaded my recording and asked it to extend the song. I didn’t even write a prompt. After listening to my recording, it made up its own prompt based on the music:

“A smooth jazz instrumental piece in a medium tempo featuring a prominent saxophone melody”.

You would think the AI would include the exact tempo for this song extension. “Medium tempo” is pretty broad, and “smooth jazz” is something entirely different stylistically. It went on:

“The song is in a major key, likely C major, with a relaxed and improvisational feel”.

True, but again, very vague. There are a bunch of chords, how is it going to decide which ones to play?

“The primary instruments are saxophone, piano, bass, and drums”.

All true.

“The saxophone carries the main melodic lines, often with a warm, slightly reverbed tone”.

AI thinks I have a warm tone. Very nice.

“The piano provides harmonic support with jazz chords and occasional fills. The bass plays a walking bassline providing a steady rhythmic and harmonic foundation. The drums maintain a light swinging rhythm with brushes on the snare and ride cymbal”.

This part is not accurate. It’s clearly not a swing rhythm, it’s more of a calypso-inspired Latin groove.

“The production is clean with a balanced mix that allows each instrument to be clearly heard. There are no vocals present. The song structure is typical of jazz with a head-solo-head format, though only the head is present in this excerpt. The melody is lyrical and flowing with a bluesy inflection at times”.

That last part is accurate, but there’s no mention of how many measures the form is. After reading that prompt, I was sure there’s no way this thing is going to generate an extension that matches the original recording. But I was wrong.

What it produced is uncanny. AI still leaves audible artifacts in the music, and if you listen very carefully, you can hear all sorts of weird things going on. But on the surface, the way 99.9% of normal people listen to music, this sounds real. Not only does it sound real, but the saxophone is playing in a style very similar to the way I play.

There is so much craziness going on with this recording. Not only did the AI continue playing a convincing solo in my style, it kept the form and chord changes to St. Thomas without any prompting. It then goes on to add a piano solo.

After the piano solo, AI Jay comes back for another improvised chorus that sounds pretty darn good, if I do say so myself.

And here is where things get pretty scary, it then plays the melody to St. Thomas. Exactly. It even adds convincing embellishments. And if that wasn’t crazy enough, the ending almost made me lose it.

It just knew how to end St. Thomas. Almost. It ends with a blues line, which is something I might do, because my favorite musicians do that a lot. Isn’t it interesting that AI can identify blues? It’s in the prompt. It heard me playing and decided there was an occasional blues feel and therefore added that into the solos as well as the ending.

Honestly, I don’t know what this means for music or musicians. It doesn’t serve any practical purpose for me, I wouldn’t use this recording in any professional way, but I’m sure a lot of other people would and are.

Why hire a writer when you can use AI? Why hire a graphic designer when you could use AI? Why hire mixing and mastering engineers? Why hire musicians to record original music or even covers when you can use AI for a negligible cost?

Please share your thoughts on this. If anything, this AI business should encourage humans to spend more time learning to play instruments, not to start a career, but to experience the unique benefits that we can only get from playing music.

Click here to read about how long it actually takes to learn the saxophone. When done right, it doesn’t take as long as you might think.

Backing Tracks (by humans) from BetterTrax