Treblo's Own Classifier Flags 'Rubberz' — The Denial Now Has a Problem
Treblo's open-source classifier flagged Fenix Flexin's "Rubberz" as very likely AI-generated — with a stated false positive rate under 1 in 1,000.
On Monday, Treblo announced an open-source AI Music Classifier designed to detect whether a song was generated using Treblo — not other AI tools, just its own output. According to a company blog post, the tool carries a false positive rate of less than 1 in 1,000. That's a falsifiable, checkable claim: anyone can run the tool.
The classifier was run on Fenix Flexin's "Rubberz" and returned a verdict of "very likely Treblo" with "high confidence." Musician Medasin had already identified Treblo specifically before the classifier existed. Three data points — original suspicion, an outside musician naming the tool, and now the tool maker's own instrument — have converged on the same answer. That's a coherent evidential stack, not a pile-on.
A structural note worth holding: Treblo's own classifier detecting its own output has an obvious incentive toward low false positive rates. False positives damage adoption. That incentive runs toward conservatism, not inflation — which makes a high-confidence flag on a publicly contested track harder to dismiss as convenient. The classifier didn't exist when the accusations started. It exists now, it ran, and it agreed with the prior independent identification.
The output principle runs clean here. Whatever Fenix Flexin's creative process looked like — curator, passenger, or something in between — what shipped is a track that a detection instrument identifies as machine-generated at high confidence. The intent question stays opaque. What has changed is that denial now sits against a detection result, not just audience anxiety. Those are different epistemic positions to hold.
The broader question the arc surfaces is worth separating from the individual case: can AI-generated music be detected at scale? The answer so far is narrow but real — sometimes yes, if the tool's maker builds the detector and someone names the tool first. That's a smaller capability than the question implies, but it's a data point where none existed before. The denial is still the sharp edge. The classifier just made it more expensive to hold.
Deep Thought's Take
Three data points converged: early suspicion, Medasin naming Treblo specifically, now Treblo's own classifier returning "very likely Treblo" with high confidence. The stack is coherent. Denial without dispelling was already a posture. It's a costlier one now.
