Contents
You may have heard, or at least heard of, the Velvet Sundown. The band makes vaguely psychedelic ’60s and ’70s protest anthems; their first album, Floating on Echoes, debuted in June, and a few weeks later they released another record, Dust in Silence. They have more than 1 million monthly listeners on Spotify, where their music is being pushed via “Discover Weekly” playlists, and their most popular track, “Dust on the Wind,” hit Spotify’s Viral 50 chart in the U.K., Sweden, and Norway. Though popular, it’s fair to say that the music is derivative, and that’s because it quite literally is: The band — with Gabe Farrow, Lenny West, Milo Rains, and Orion Del Mar as its “members” — and its music are all a product of AI.
Their success is raising the question of whether generative-AI music has improved so much that listeners can no longer hear the difference. Certainly, AI music can meet our expectations of song structure. The Velvet Sundown’s melodies follow chord progressions, and their songs have rhythm sections that follow along with the bass and drums. The backing track gives space to the vocalist, and when the vocalist drops out, the instruments come back in and you get a call-and-response. More often than not, it genuinely seems as if people are playing this music.
And yet they are not. Time will tell how prevalent AI-generated music will become — and, more critically, whether audiences will even care about a song’s origins if the track is a bona fide bop. But in this moment, as we try to understand this emergent technological and its implications, it seems important to be able to identify when a piece of music is made by AI. As my co-host, Nate Sloan, and I discussed in a recent episode of Switched on Pop, there are five clear tells — for now.
1. Bad Rhymes
In AI lyrics, something is just a little bit off. There is a constant barrage of single syllable, perfect-end rhymes: The drums roll slow / Tell me, brother, where do we go? The rhyming is basic and exact. In songs written by humans, things are more likely to almost match up, but not quite. They are just unpredictable enough. In addition to the robotic exactness of the lyrics’ sound in AI-written songs, they are usually devoid of meaning. Velvet Sundown songs purport to be protest anthems, but when you look at them, there’s no substance and no specificity. They talk about making enemies pay, and then immediately contradict themselves by talking about praying for peace. There’s no cohesion at all.
2. Excessive ‘Borrowing’
Often you’ll hear an AI-generated song and you’ll say, “Dang, that feels really familiar.” That’s because it is. AI music, to put it kindly, is fond of “borrowing” from existing popular songs. The Velvet Sundown is not inspired by bands like Pink Floyd and Buffalo Springfield, they have literally ripped them off. Listen to their music and you hear exact phrases pastiched together to make a new Frankenstein’s monster. If you hear something you’ve heard before and it keeps happening, chances are good it’s AI.
Click here to preview your posts with PRO themes ››
3. Inconsistency
Most bands have a recognizable sound from song to song, even with variations. If every single song sounds like there’s a different singer, it might be AI. From one Velvet Sundown song to another, you get a voice that’s much more nasally, one with a much purer tone. Even the guitars sound kind of different. There’s no consistency of identity. Bands can have guests, people playing different parts, but if they claim to have the same lineup and swing wildly in tone, you may have stumbled onto AI music.
4. Strange Song Structure
Bizarre song structure used to be a more conspicuous tell: Sometimes you’d have a chorus No. 1 that wouldn’t sound like chorus No. 2. That’s gotten better, but you can still find ways in which an AI is not aware of the odd choices it makes. You’ll hear a song that trails off into the end unexpectedly; it sounds like the song just quits. Occasionally, you may also get unexpected interludes or double intros. Velvet Sundown, of course, has all of these things.
5. Ghost Instruments
If you listen closely, on many AI songs, you’ll discover sounds that are kind of just frankly indistinguishable ghosts of instruments that never fully existed on the finished product. There will be traces of horns in the background, blurry sounds layered on top of the expected instruments. They’re often in a background layer when things get a little bit fuller in the mix. It doesn’t sound purposeful — it sounds incongruous and uncanny.
Now, a quiz: Can you tell which of these songs are made by humans and which came from AI?
Listen to each track and see if you can determine its provenance.
"Untitled"
Human
A.I.
Right. This is from the "Nobody in the Computer" YouTube channel. It’s Bill Evans–style jazz created by training on the pianist’s recordings.
Not quite. This is from the "Nobody in the Computer" YouTube channel. It’s Bill Evans–style jazz, created by training on the pianist’s recordings. "Phil Wildo"
Human
A.I.
Yep, robots. From the AI crude-humor novelty album Fool’s Golden Oldies Volume Four.
Nope, AI. From the AI crude-humor novelty album Fool’s Golden Oldies Volume Four. "Whispers of Chaos"
Human
A.I.
AI again. You might noticed this one sounds quite a lot like Mungo Jerry’s "In the Summertime." Excessive borrowing, anyone?
Close, but this is AI. This one sounds quite a lot like Mungo Jerry’s "In the Summertime." Excessive borrowing, anyone? "Echoes of Twilight"
Human
A.I.
AI indeed. Prompted as a "1970s folk song," we get a Simon & Garfunkel ripoff.
It’s AI. Prompted as a "1970s folk song," we get a Simon & Garfunkel ripoff. "Mirage Woman II"
Human
A.I.
Correct. A northern African guitar-style jam made by the machine.
It’s A.I. A northern African guitar-style jam made by the machine. "Bar Fight"
Human
A.I.
That’s right, AI this final time, too. This track is modeled on the Wu-Tang Clan.
Despite your faith in humanity, you are incorrect. It’s AI this final time, too. This track is modeled on the Wu-Tang Clan.
