Technology can transform music. Cher distorted her vocals on “Believe” with the help of Auto-Tune. Phil Collins manufactured his iconic drum break on “In the Air Tonight” with the help of a Roland CR-78 drum machine. And who hasn’t used a sampler to cut, loop, and repitch Doug E. Fresh and Slick Rick’s 1985 hip-hop classic “La Di Da Di”? (The song is sampled or referenced by Snoop Dogg, Miley Cyrus, and Beyoncé, among others.)
We know humans can make music with technology, but can tech do it alone? That’s the promise of AI music generators such as Suno and Udio. They work just like ChatGPT or DALL-E. Users type in what they want, but instead of getting back text or images, these tools produce entire songs.
We’ll get into how they actually perform in a moment, but, regardless, it’s important to know that the music industry feels threatened: On June 24, major music labels Universal, Sony, and Warner, along with the Recording Industry Association of America, the main trade association for the music industry, filed a lawsuit against Suno and Udio for allegedly violating federal copyright law. The plaintiffs claim the AI companies illegally scraped and trained their models on copyrighted material and output music that cribs from the originals.
Kristelia Garcia, an intellectual property professor at Georgetown University Law Center, said the suit is similar to those we’ve already seen from authors and artists alleging AI copyright infringement. She says the plaintiffs in this case face a real challenge in connecting individual pieces of music to specific outputs of the AI tools.
“I think the plaintiffs know that AI composition is here, and could, in fact, be very useful to their industry,” Garcia said. “I'd predict some kind of settlement in which defendants retroactively license, or offer equity or something equivalent to the record labels.”
That’s the legal battle, but let’s get to the beats: Are these music generators clear rip-off machines? And can they produce music that won’t make my ears bleed?
When I first signed up for the free version of Suno, which gave me a set number of songs before I had to pay, I typed a prompt in: “Make me a midwestern emo song about writing a newsletter for all of your homies while your dog barks in the background.”
Suno did a decent job at capturing the genre, especially the whiny timbre of emo vocals. The lyrics fit my description but weren’t very coherent. Here’s the chorus:
Sending love through pages
Every word I've crafted down
Dog's bark echoes stages
Of our hearts and small town
NEWSLETTERS AND BARKING SOUNDS AUDIO
This browser does not support the video element.
Then, Suno made me a political masterpiece. I asked for a “trap rap track with an 808 beat and lyrics about Joe Biden.” Now, take a listen to what it made, which it named “Biden’s Bounce.”
Sleepy Joe in the spot, talent on display,
Folks talkin' slow, but he quick in the relay,
Oval Office boss, drinkin' black coffee sway,
Age just a number, every day a play.
This browser does not support the video element.
If the Biden campaign wants to license this banger, you know where to find us.
Fairly pleased with Suno’s performance thus far, and glad it didn’t shy away from political commentary, I decided to test whether it would make me songs in the style of a specific artist. When I asked it for a hip-hop song about heartbreak in the style of Ms. Lauryn Hill, it told me it couldn’t honor my request because it contained an artist name. So, I dropped the “Ms.”— sorry, Ms. Hill — and it complied. The same thing happened when I requested a folk song in the style of Bon Iver. Bon Iver didn’t work but the band’s lead singer Justin Vernon did. Concerning as that may be from a brand safety perspective, neither song was any good — nor did they sound anything like the artist I requested.
This browser does not support the video element.
So, musically, I’d give Suno a C-grade, with notably higher marks for the “Biden Bounce.” It’s not anything I’d listen to, but it’s also not anything I’d immediately identify as AI-generated.
The strangest quirk about Suno came in the lyrics: The word “echoes” appeared in three of the five songs I generated. When I closed Suno and opened Udio, I gasped. The recommended prompt to try? “A song about the echo of a lost civilization.” I followed that suggestion and produced an almost-unlistenable dream pop song.
This browser does not support the video element.
I fed Udio my prompt for a midwestern emo song about writing a newsletter while your dog barks. It failed with genre, producing an indie rock song with odd vocals. But the lyrics and music were an improvement over Suno, better fitting what I asked of it.
This browser does not support the video element.
To finish my experiment, I repeated another prompt, asking Udio to make me a Lauryn Hill song. “Heart Forsaken,” the song it produced, was a much closer sonic approximation of an artist’s musical style than anything else I was able to produce. It didn’t sound like her, but the underlying beat and lyrical stylings weren’t terribly far off.
This browser does not support the video element.
Perhaps these AI-generated songs could be background music in a store, strung together on a long YouTube video called “Music for Studying,” or some other trivial purpose. But the record industry — until further notice — needn’t freak out.
There’s no AI-generated Taylor Swift or Olivia Rodrigo coming to threaten its current stars. Whether it’s violating copyright, however, the courts will decide.