Treblo's own detector flagged a Hot 100 hit as its output. That is the real power shift

For most of the summer, the question hanging over Fenix Flexin's "Rubberz" was a matter of taste and suspicion: does that bright, Morrissey-ish voice really belong to a Shoreline Mafia rapper? At the start of August it became a matter of software. The company whose app the song was allegedly made with released a detector for its own output, and the detector said yes.

What happened

"Rubberz" came out in June. By early August, WIRED reported it sitting at No. 58 on the Billboard Hot 100, with a video at 7 million views. Listeners had questioned it from the start, because it is a synth-pop song with a British accent from an artist known for something very different, and the voice doesn't sound like his. Fenix and the credited producer, Purps on the Beat, denied using AI. In one reply to a fan he put the differences down to auto-tune, reverb and a put-on UK accent.

None of the well-known AI detectors flagged the track, WIRED says, so the argument stayed open. Then producer Medasin posted a breakdown arguing the song was made with Treblo, a smaller generator that until two days before the song's release was called Sonauto. His evidence included a filename ending in "Sonauto.mp3" glimpsed in one of Fenix's old Instagram Stories, and Treblo generations he prompted himself that came back with strikingly similar lyrics. WIRED repeated the test and got near-identical rhymes.

Fenix posted what he said was the project file. Pitchfork reports that commenters said the stems looked like the output of a stem separator, which can split any finished song into parts. He deleted the clips.

The detector

Treblo then published a page called "Was This Made With Treblo?". Its accompanying post, as quoted by WIRED, says it is tuned to catch songs made with Treblo "with little to no modification", and frames it as something listeners should have. It doesn't detect other generators. Run "Rubberz" through it and it returns "very likely Treblo"; a track from Tyga, Fenix's labelmate, came back "likely Treblo".

Treblo says the false-positive rate is under 1 in 10,000, and it released the code under an open-source licence; WIRED notes the detector has not been independently reviewed. The Verge reported the rate and later corrected its own first report of the rate, which said 1 in 1,000. Treblo's Ryan Tremblay told The Verge the score indicated the released audio was generated almost entirely by Treblo's model, and claimed that would make it the first known AI-generated song to reach the Hot 100. That is the company's claim, and it is a flattering one for the company. To WIRED he said Treblo hadn't talked to Fenix or Tyga and didn't know the song existed until it charted.

Within days, Fenix changed his account. In an Instagram comment reported by Complex and in Pitchfork's update, he said he had never claimed not to use AI, only that it played no part in the recording process, and that he didn't know what the app was until after the mix. He added that artists should keep up with the times.

Why this matters to artists

The obvious reading is "AI song caught". The more useful one is about who did the catching.

Third-party detectors try to spot AI music in general, and they missed this one. Treblo's detector only has to recognise one model's fingerprints, and the company that owns the model controls whether it exists, how sensitive it is, and when it is switched on. Treblo published its detector after a summer of free publicity, in the middle of a public argument, and the result helps Treblo's marketing whatever it does to the artist. In practice an output-specific detector works like a watermark that only the vendor can read and only the vendor decides to reveal.

That cuts both ways. A tool like this could protect listeners and human artists from passing-off. It also means anyone who generates on a platform has handed that platform a record it can later use against them, or for itself. None of the coverage we read says whether Treblo's terms warned users that their tracks could one day be publicly identified this way, and Treblo's site blocked us when we tried to check.

Artists are not all in Fenix's position, either. One songwriter wrote on X that he writes every lyric himself but uses AI to sing them because his own voice and delivery aren't strong enough, sometimes guiding it with short recordings of his own. A yes-or-no detector result has no room for that. Neither does the word "snitch", which is how one indie-label account on X framed the story, noting that it landed the same week Spotify signed Merlin for its upcoming AI covers and remix tool (Merlin. The same week, the industry was licensing AI in through the front door and hunting for it at the back.

Disclosure systems do exist. Spotify said in September 2025 it would support an industry standard for AI disclosures in credits, developed through DDEX, covering AI vocals, instrumentation and post-production. But it depends on artists and distributors filling it in, and Spotify itself says the absence of a credit doesn't mean AI wasn't used. "Rubberz" shows what happens in that gap: the disclosure came from a vendor, then a denial, then a partial admission.

What we think

Disclosure on the artist's own terms beats being outed by a vendor, and artists should treat every generator as a potential witness. If you use a generation tool, assume its maker can identify the output now or later, and decide what you disclose before someone else does it for you. Detectors that only recognise their own model's output shouldn't be read as neutral evidence, because the party running them profits from the verdict. Labs that ship one should publish their accuracy testing, tell users in their terms that output can be identified, and give artists a route to respond before results go public.

The bigger failure is that the norm for AI disclosure in pop is being set by a scandal rather than a standard. Platforms have now started labelling AI music, each in its own way, and that still doesn't give artists a clear place to say how they used it. A credit should separate a song generated end to end from one where AI assisted a human performance or arrangement. Assistive tools, including Veena, the browser DAW with an AI co-producer, sit in a different place from full generation, and an honest credit line should be able to say which one a track used. Until that exists, the most reliable disclosure mechanism in the business is a company deciding it's good for marketing.

Sources

• Did an AI Music App Just Snitch on the Song of the Summer?, WIRED
• Sure seems like Fenix Flexin used AI music generator Treblo, The Verge
• So Is Fenix Flexin's "Rubberz" AI or What?, Pitchfork
• Fenix Flexin Does U-Turn, Says He Did Use AI to Make Viral Hit "Rubberz", Complex
• Spotify Strengthens AI Protections for Artists, Songwriters, and Producers, Spotify Newsroom
• Spotify and Merlin Announce Licensing Agreement for Fan-Made Covers and Remixes, Spotify Newsroom
• Post on writing lyrics and using AI to perform them, X
• Post on the Treblo detector and the Spotify-Merlin remix deal, X