It was the debut that few artists dream of. One month after releasing their first songs in June, psychedelic rock band Velvet Sundown <\/strong>had raked in more than 900,000 listeners on Spotify alone.<\/p>\n Just one problem: the Velvet Sundown wasn’t real. The ‘band’ was AI-generated, though that only became confirmed by journalists after people questioned its appearance on their Discover Weekly playlists. Currently, major streamers don’t let listeners know if an act is AI-generated or fake.<\/p>\n The thing is, it wasn’t exactly a slam-dunk of a dupe. The band was clearly AI-generated if you paid any attention. First, the images.<\/p>\n While the four-piece had the right number of fingers, their skin was bizarrely glossy, backdrops too generic. Nothing quite looks right if you stare for long, and members looked slightly different in each shot.<\/p>\n Another tell-tale sign? The band’s members, listed off in its bio, had no signs of existence; the band’s social presence was light despite their popularity, and they had seemingly never performed live, either.<\/p>\n And as for the music itself? It’s not overtly robotic \u2014 definitely derivative of ’70s folk greats, but nothing offensive. But while the instrumentation never shifted much between songs, the singer’s voice did considerably.<\/p>\n Loading<\/span><\/span><\/p>\n So how does a random, relatively obvious AI act that makes run-of-the-mill music get the kind of streaming numbers that most new acts would kill for?<\/p>\n No, the answer isn’t bots. (While people still use AI and streaming farms to generate royalty payments, Spotify, Apple Music, YouTube and other streaming services have clamped down on the practice, making it harder to pull off \u2014 especially at the scale of Velvet Sundown’s streams.)<\/p>\n According to Liz Pelly<\/strong>, a music journalist and author of Mood Machine: The Rise of Spotify and the Costs of the Perfect Playlist<\/em>, Velvet Sundown’s genericness is the secret to its streaming success.<\/p>\n “I think a lot of streaming companies or AI companies are not counting on people just stumbling across this [kind of] material and thinking it’s great,” she says.<\/p>\n “They’re counting more on people not noticing it in the background, in an algorithmic playlist recommendation or something like that\u2026 they’re trying to game the system.”<\/p>\n Liz says Velvet Sundown likely reached real listeners after its tracks were “juiced” algorithmically by appearing on popular user-generated playlists \u2014 potentially paid for.<\/p>\n As of writing, Velvet Sundown appears on several popular playlists, including ‘Supernatural Soundtrack’ and ‘Vietnam War Music’. The latter has more than half a million saves and features a heap of Velvet Sundown songs among its extensive list of genuine ’60s and ’70s music.<\/p>\n Something that doesn’t draw attention to itself \u2014 say, a pastiche of the great bands it’s playlisted alongside, feasibly a lesser-known track by a contemporary \u2014 will likely generate streams, as people passively listen.<\/p>\n “That gets the tracks on some Discover Weeklys and probably some other personalised offerings,” she says. “A snowball effect happens from there. “<\/p>\n After Velvet Sundown was outed as AI-generated, its streaming numbers have sliced in half.<\/p>\n But that’s still 400,000 monthly Spotify listeners listening to the “synthetic music project”, as it’s now described in its bio, having dropped the bandmate’s fake names.<\/p>\n While the most well-known AI-generated act for now, Velvet Sundown is just one of hundreds of thousands of fake artists on streaming services.<\/p>\n In recent years, AI music generators such as Suno, Udio and Boomy offer paid users commercial use of their creations, meaning they can upload their tracks to streaming and pocket the royalties.<\/p>\n While major streaming services tend not to reveal numbers, French streaming service Deezer <\/strong>recently announced<\/a> that 18 per cent of its daily uploads, or roughly 20,000 new tracks a day, are AI-generated. As a result, Deezer is pulling AI-generated music from its algorithmic recommendations.<\/p>\n But Liz says that AI-generated artists make up a small sliver of fake acts on streamers.<\/p>\n Loading<\/span><\/span><\/p>\n Outlined in great detail in Mood Machine<\/em>, Liz identifies the majority of fake acts on streamers as “ghost artists” \u2014 intentionally generic, stock-like music uploaded under fake artist names, some of which is commissioned by streamers themselves.<\/p>\n “I was writing about this phenomenon of production companies that had licensing agreements with Spotify, hiring session musicians and production music makers to make music in bulk,” she says.<\/p>\n Spotify calls this program Perfect Fit Content<\/strong>, where they commission music inoffensive enough to slip into playlists largely unnoticed under generic artist names. (Jazz, ambient and electronic-based playlists, like study, sleep and other “functional” playlists are home to a lot of PFC).<\/p>\n And because Spotify \u2014 and other streamers who have similar licensing agreements \u2014 own these songs, they don’t have to pay royalties on them.<\/p>\n Not all ghost artists are created by streaming companies for their own means, though. Independent artists, licensing companies and other external parties have also tried to game the system by making generic music geared towards algorithmic recommendations.<\/p>\n And it can work: Last year, The Guardian<\/em><\/a> reported that a virtually unknown composer Johar R\u00f6hr<\/strong> was Sweden’s top-streamed Spotify artist, with 700 songs uploaded under 650 different names collectively earning him 15 billion streams and potentially millions in royalties.<\/p>\n With a single song stream generating anywhere between $0.00113 and $0.012 across popular streaming sites (before an artist’s label and management take a cut), it can take a million listens to earn a thousand dollars.<\/p>\n In short, most artists aren’t making massive amounts from streaming.<\/p>\n And ghost artists are effectively cutting into the opportunities to gain audiences via playlisting or algorithmic recommendations.<\/p>\n AI-generated artists are essentially extra salt to the wound, as it’s unclear where many of them take the data used to create these songs from.<\/p>\n “The big issue is consent,” says Liz. “Did artists consent to their work being used as training data? Are artists forced to compete in a marketplace against tracks that were generated and trained on their work, without their permission?”<\/p>\n And while some AI music companies do purport to use only licensed music, Liz argues that these AI-generated acts still squeeze the limited royalty pool, leaving real musicians worse off.<\/p>\n She’s far from alone with her concerns. Record labels have repeatedly sued <\/a>AI-generated music using their artists’ voices \u2014 most famously, a fake Drake and the Weeknd collab in 2023 was taken down <\/a>after going viral.\u00a0<\/p>\nUnderstanding Velvet Sundown and AI-music’s rise<\/h3>\n
<\/div>\n<\/div>
<\/div>\n<\/div>
The tip of the iceberg<\/h3>\n
<\/div>\n<\/div>
Why does this matter?<\/h3>\n
<\/div>\n<\/div>