When the new “ethically trained” artificial-intelligence-powered music generator Jen dropped a couple of weeks ago, it was at a fairly auspicious time: The Recording Industry Association of America had just sued Udio and Suno for copyright infringement, alleging that the programs were trained on copyrighted material without the creators’ permission. By contrast, Jen claimed to have licensed more than 40 training catalogs and promised to vet everything that went into, and came out of, its system, to make sure it didn’t violate copyright. As other AI music generators were garnering less-than-flattering headlines, Jen was positioned to be an alternative.
Ethics aside, whether or not the music Jen generates stands up is debatable. Tracks generated by the program don’t have lyrics, and while its AI was trained on licensed material—the list of which hasn’t been released—it doesn’t seem to be anything you might have heard on the radio. You can’t ask Jen to make a Willie Nelson song because it doesn’t actually know what Willie Nelson sounds like. (Our attempt to do so produced something that sounded more like trip-hop than anything else.) Even if you sub out Nelson for “outlaw country,” a genre that has existed for half a century and has influenced countless landmark LPs, you still just end up with something blandly “country,” more easy listening than honky-tonk.
To dig deeper, WIRED asked five professional musicians to test Jen’s skills: John Heywood, a bassist who’s best known for backing indie rock act Alex G; Wye Oak and Flock of Dimes’ Jenn Wasner; Shana Cleveland, founder of surf-noir act La Luz; Steve Reidell, of The Hood Internet and Air Credits; and Allen Blickle, a two-time Emmy-nominated composer and sound designer who has worked on projects for Netflix, Disney, and Apple Music, among others. All five found the program easy to use but inherently uninspiring.
Wasner, who says she’s open to the idea of AI as a “tool to help generate ideas,” has seen AI programs used in the studio or in a band’s writing process before. In her time with Jen, though, she struggled to make anything she could take to heart. “Everything it made seemed like it came from an uncanny valley situation, and while it was fascinating to listen to, it all just felt like a trick, like ‘Oh, I can put trap high hats on a bluegrass track,’” she explains. “There was never a point when I thought, ‘That’s a cool idea.’ I always thought, ‘I could have come up with something cooler on my own.’”
Admittedly, our testing artists did push Jen beyond the boundaries of what a “normal” person might ask in a query, veering more toward a “record store clerk” level of familiarity with recorded sound. Cleveland, for instance, failed to get anything good out of a query for “mid-tempo California garage rock influenced by ’70s Indonesian pop,” while Heywood expressed dismay that Jen didn’t seem to recognize his request for “city pop,” a type of Japanese music that came to prominence in the mid-’70s and has seen a minor resurgence in popularity in recent years. But to Heywood, that kind of breadth of music is necessary, especially as a musician.
“A lot of musicians or producers, when they ask something of each other, they’ll use bands and other artists as a reference point, like, ‘We’re going to go for a Prince type of sound,’ or, ‘Let’s add some Clavinet like Stevie Wonder,’” Heywood explains. With Jen’s lack of understanding of both existing recording artists and even some fairly common genres and instruments, it makes it hard to really land on something specific.
“I kept trying to coax some warmth out of it, like vinyl hiss or saturation or something lo-fi or vintage sounding, but everything it made had the same kind of hi-fi, video-game-menu-screen-type sound to it,” Heywood says. “They even give you ‘lo-fi’ as a prompt suggestion, but that didn’t seem to make much of an impact. If you’re trying to get a certain sound, like ’80s funk, the closest you’re able to get is something that sounds more like Daft Punk.”
Every electric guitar sound that WIRED and the testers generated sounded almost too clean, and it was virtually impossible to get it to produce a track that wasn’t in a 4/4 time signature unless you used the word “waltz” in the prompt.
Some of this, says Jen cofounder Shara Senderoff, is to be expected. The tool is in its alpha phase, and the 10-second and 45-second tracks it generates are “meant to inspire and provide a starting point for creativity, not necessarily a final product,” she says. New capabilities are coming, and because Jen was trained using a limited data set, it has room to grow and “will expand significantly in the beta phase,” Senderoff adds.
Everything Jen made under the guise of rock music, Heywood says, was akin to “the clip art version” of the genre. Cleveland was able to coax out some songs that sounded “like they could be used in a car commercial” or that were “getting into Black Keys territory,” but says more than anything, she felt like all Jen’s musical suggestions were just plain hokey.
“It felt like the kind of music I’d make if I were messing around with my friends, joking about the cliches of other genres,” she says. “I could see some of the songs on a super bad Netflix dating show, but nothing I made felt like a threat to me personally.”
But what about everyone who makes the tracks you might hear on a Netflix dating show? Could Jen be a threat to their jobs? According to Blickle, almost certainly.
“If you’re a producer with a small budget and you’re just trying to get your content out, now you can say, ‘I’m not even going to pay a designer or an animator. I can just use an image generator,’” he says. “The same thing is true for a music budget. If they can pay nothing for something that was going to cost them $2,000, then great, someone will think that’s $2,000 in their pockets.”
Apps like Jen don’t yet give creators the ability to set crescendo points or add stingers to their tracks, which would make them ready-made for scoring, but one can only expect that it’s on the horizon.
Blickle guesses that stock music libraries, which a lot of lower-budget productions and reality TV shows use to score their product, will see an influx of material from sources like Jen. While he calls a lot of what’s on those libraries “trash,” he says that “if you’re looking for that type of music, there’s just going to be more out there. It’s going to create a feedback loop of crappy library music and I don’t see how that betters anyone’s enjoyment of life.”
It could also really muck up what’s available on streaming music sites like Spotify, where thousands of new songs are already added every day. “That’s going to multiply by two, three, four, five times, just because of people who are using AI in hopes of getting 100,000 plays on a song, or maybe a million plays on another,” Blickle explains. “They just want to release music and make a little extra money, and that’s scary.”
An onslaught of AI-created music could also generate problems for productions that choose to use AI music. While sites like YouTube and Instagram would let you post an AI-generated song, since you would in theory own that “new” track, the laws for AI-generated content vary around the world, meaning that if you’re looking to sell a TV show into other markets, you might want to be careful about how your music was created—even if it comes from an “ethical” source like Jen. US law also dictates that any work seeking copyright protection has a human author.
Twenty-five years after Napster, all five artists said the possibility of music generators like Jen—ethically sourced or not—decimating their livelihoods just felt kind of inevitable. “There are so few ways left for musicians to make money doing the thing that they do, and to have another one of them eroded and chipped away at is definitely unsettling. But I’m in a place where I feel like I’ve already grieved that loss,” Wasner says. “We’re just going to have to figure it out.”