A rap song that climbed the Billboard Hot 100 might be the first chart hit generated almost entirely by AI, and the company behind the tool just handed everyone the receipts. According to The Verge AI, the AI music startup Treblo announced on Monday an open-source classifier that flags songs made with its own model, and that classifier says Fenix Flexin’s track “Rubberz” is “very likely Treblo” with “high confidence.” The Verge AI reports the detection tool carries a false positive rate of less than 1 in 10,000.
What stands out here is the direction of the accusation. Usually artists deny AI involvement and there’s no hard way to prove it. This time the AI company itself built and released the receipts.
What actually happened
Musician Medasin suspected “Rubberz” was made with Treblo specifically, not just some generic AI tool. Treblo’s new classifier appears to back that up.
Here’s the sequence, per The Verge AI:
- Treblo released the open-source Treblo AI Music Classifier, built to detect its own model’s output (not other AI tools).
- The blog post claims a false positive rate under 1 in 10,000.
- Run against “Rubberz,” the classifier returned a very high probability score.
- Treblo CEO Ryan Tremblay told The Verge the released audio was “generated almost entirely by our model,” which “would make it the first known AI-generated song to reach the Billboard Hot 100.”
Fenix isn’t backing down. He’s still denying AI involvement and has posted clips of what he says are the project files from the recording session. The Verge AI notes those low-quality videos have done little to quiet the allegations. Medasin argued in a since-expired Instagram story, via Stereogum, that Fenix simply used an AI stem separator to fake the session. Fenix didn’t immediately respond to a request for comment.
Why this matters
The fight over AI in music has been stuck on a simple problem: proof. You could hear something and feel it was synthetic, but feeling isn’t evidence. A model-specific detector changes that math.
A few things worth sitting with:
- A detector tuned to one model is far more credible than a general “is this AI?” guess. Treblo knows its own fingerprints.
- A sub-1-in-10,000 false positive rate is strong, though the company is quick to admit it isn’t “unimpeachable.”
- Treblo is clearly enjoying the moment. Tremblay framed it as someone using the model “to find a sound that millions of people connected with,” which reads as a marketing win dressed up as a confirmation.
That last point deserves a raised eyebrow. The company confirming the story is also the company that benefits most from being credited with a Hot 100 hit. That doesn’t make the classifier wrong, but it’s worth remembering who’s holding the microphone.
What comes next
Expect this to get messier before it gets cleaner. A few things to watch:
- Other labs may follow. If model-specific detectors become standard, artists lose the “you can’t prove it” defense for that particular tool.
- The stem-separator loophole. Medasin’s theory, that someone ran real audio through AI and then claimed a live session, is going to be the go-to cover story. Detectors and defenses will keep leapfrogging each other.
- Charts and credits. If an AI-heavy track can quietly hit the Hot 100, labels, streaming platforms, and rights groups will face pressure to define disclosure rules fast.
For anyone working with generative audio, the lesson is blunt: the output can be traced back to the tool that made it, at least when that tool decides to build the detector. Transparency may soon be less of a choice and more of a default.
The full breakdown, including Tremblay’s statement and the back-and-forth with Fenix, is available at the original report from The Verge AI.