AI music platform Suno’s policy is that it does not permit the use of copyrighted material. You can upload your own tracks to remix or set your original lyrics to AI-generated music. But, it’s supposed to recognize and stop you from using other people’s songs and lyrics. Now, no system is perfect, but it turns out that Suno’s copyright filters are incredibly easy to fool.
With minimal effort and some free software, Suno will spit out AI-generated imitations of popular songs like Beyoncé‘s “Freedom,” Black Sabbath’s “Paranoid,” and Aqua’s “Barbie Girl” that are alarmingly close to the original. Most people will likely be able to tell the difference, but some could be mistaken for alternate takes or B-sides at a casual listen. What’s more, it’s possible someone could monetize these uncanny valley covers by exporting them and uploading them to streaming services. Suno declined to comment for this story.
Making these covers requires using Suno Studio, available on the company’s $24-a-month Premier Plan. Rather than prompting a whole song with text, Suno Studio lets you upload a track to edit or cover. It’s likely to catch and reject a well-known hit with no tweaks. But using a basic free tool like Audacity to slow down a track to half-speed or speed it up to twice normal will often bypass the filter, and adding a burst of white noise to the start and end seems to basically guarantee success. You can restore the original speed and cut the white noise in Suno Studio, and the copyrighted song becomes the seed for new AI music.
If you generate a cover of the imported audio without any style transfers, Suno basically spits out the original instrumental arrangement with very minimal tweaks to the sound palette if you’re using model 4.5 or 4.5+. Model v5 is a bit more aggressive in taking liberties with the source material, adding chugging guitar and galloping piano to “Freedom” and turning the Dead Kennedys’ “California Über Alles” into a fiddle-driven jig.
Suno lets you add vocals by generating lyrics or typing words into a box, and once again, it’s supposed to block anything copyrighted. If you copy and paste the official lyrics for a song from Genius, Suno will flag them and spit out gibberish vocals. But extremely minor changes can bypass this filter as well.
I was able to trick Suno Studio by tweaking the spelling of a handful of words in “Freedom” — changing “rain on this bitter love” to “reign on” and “tell the sweet I’m new” to “tell the suite” — and beyond the first verse and chorus, I didn’t even need to do that. The voice closely mimics the original recording, summoning slightly off-brand renditions of Ozzy or Beyoncé.
Indie artists might not even be afforded that level of protection. One of my own songs cleared the copyright filter while I was testing v5 of the company’s model. I was also able to get tracks by singer-songwriter Matt Wilson, Charles Bissell’s “Car Colors,” and experimental artist Claire Rousay by Suno’s copyright detection system without any changes at all. Artists on smaller labels or self-distributing through Bandcamp or services like DistroKid are most likely to slip through the cracks; DistroKid and CD Baby declined to comment.
The results of these AI covers fall firmly in the uncanny valley. The songs they’re covering are unmistakable: the riff from “Paranoid” remains identifiable and “Freedom” is obviously “Freedom” from the moment the marching snare hits kick in. But there is a lifelessness to them. Even if AI Ozzy is alarmingly accurate-sounding, it lacks nuance and dynamics, leading it to feel like an imitation of a human, rather than the real thing.
The instrumentals similarly discard any interesting artistic choices the originals make, or clone them in flat imitations. A non-jig “California Über Alles” cover has most of its rough edges sanded down so it sounds like a wedding band version of the original. Pink Floyd’s “Another Brick in the Wall” goes from an experiment in doom disco to just vacuous dancefloor filler. And, while it kind of nails David Gilmour’s guitar tone, it does away with any sense of phrasing or progression, turning the solo into just a mindless stream of notes.
Creating unauthorized covers violates both the stated purpose of Suno, and the terms of service. Moreover, Suno only appears to scan tracks on upload; it doesn’t seem to recheck outputs for potential infringement, or rescan tracks before exporting them. The path to monetizing Suno-created covers is simple from there. AI slopmongers could upload them through a distribution service like DistroKid and profit from other people’s songs without paying the typical royalties a cover would give the original composer. And independent artists seem to be the most vulnerable.
Folk artist Murphy Campbell discovered this recently when someone uploaded what seem to be AI covers of songs she posted on YouTube to her Spotify profile. (It’s not clear what system they were generated through.) Shortly afterwards, distributor Vydia filed copyright claims against her YouTube videos and began collecting royalties on them. And to highlight just how broken the whole system is, the songs which Vydia successfully filed copyright claims for are all in the public domain. Spotify eventually removed the AI covers, and Vydia has rescinded its copyright claims, but that only happened following a social media campaign by Campbell. Vydia says the two incidents are separate and it is not associated with the AI covers of Campbell’s work.
AI fakes are a problem for other artists too. Experimental composer William Basinski and indie rock group King Gizzard and The Lizard Wizard have had imitations slip through multiple filters and reach streaming platforms like Spotify. Sometimes, these fake songs can siphon up views straight from the artist’s own page. In a system where payouts can already be brutally low — Spotify requires a minimum of 1,000 streams to get paid — less famous musicians are hit hardest.
Suno is only one cog in a clearly broken system.
Services like Deezer, Qobuz, and Spotify have taken measures to combat spammy AI and impersonators. Spotify spokesperson Chris Macowski told The Verge that the company “takes protecting artists’ rights seriously, and approaches it from multiple angles. That includes safeguards to help prevent unauthorized content from being uploaded in the first place, along with systems that can identify duplicate or highly similar tracks. Those systems are backed by human review to make sure we’re getting it right.” But no system is perfect, and keeping up with a flood of AI slop enabled by platforms like Suno poses a challenge.
Macowski acknowledged the technical difficulties involved, saying, “It’s an area we’re continuing to invest in and evolve, especially as new technologies emerge.”
Suno is only one cog in a clearly broken system. But it’s one artists have particularly little recourse to fight. Bands can contact Spotify and have AI fakes removed from their profile. It’s harder to tell how those fakes are generated, and if they’re the result of Suno’s filters failing. And so far, Suno’s response is silence.
Read the full article here
