wired.com web signal

HumanStandard Now Flags Hit 'Rubberz' as Made by Treblo AI

TL;DR

  • 'Rubberz,' credited to Fenix Flexin and Purps on the Beat, is a summer hit now accused of being generated on AI music app Treblo.
  • Detection platform HumanStandard initially cleared the track as human-made, then reversed after a tip pointed founder Rasha Rahman to Treblo, formerly Sonauto.
  • HumanStandard says it now flags Treblo tracks at roughly 90% accuracy after just 1,000 training samples, versus the 10,000 needed for Suno detection.

The most interesting thing about the fight over whether 'Rubberz,' the summer's stickiest rap earworm credited to Fenix Flexin and Purps on the Beat, was made with an AI music generator is not the accusation itself. It's how close it came to slipping through undetected, and what pulled it back into the light.

According to Wired's reporting, an AI detection platform called HumanStandard initially cleared 'Rubberz' as human-made. Its artifact detection model returned a near-zero AI probability, and its hybrid voice model flagged the track as only partially AI. HumanStandard is built on a zero-false-positives premise, so in doubt it erred toward human. Then founder Rasha Rahman got a tip that the song had been made on an AI song generator he had never heard of, called Treblo. Treblo, formerly named Sonauto, had rebranded roughly two days before 'Rubberz' came out, which is a good part of why it sat outside the training distribution of most existing detectors. It is also why 'Sonauto' turned up in a song file name Fenix himself had posted, seized on by producer Medasin in a viral breakdown video that Stereogum recapped.

Once Rahman had a name to chase, retraining was quick. HumanStandard reports it now detects Treblo tracks at roughly 90% accuracy from just 1,000 training samples, a tenth of what Suno detection required when Rahman first tackled that platform. The reversal is the useful part of the story, because it maps the arms race in a single case study. A new generator ships or rebrands, detectors need a tip and a small sample set to catch up, and in the gap between those two events a track can climb the charts.

The honest caveat is that both Fenix Flexin and Purps have denied the song is AI-made, and HumanStandard's flip only came after a tip rather than through blind analysis, so take the specifics as reported not as settled. What the reporting doesn't give you is what streaming services, labels, or chart bodies intend to do with a track sitting in that grey zone, or how Treblo itself views a would-be song of the summer being made on its platform without disclosure. The people best placed to benefit from the next round of this are the detection startups now sitting on a live commercial pitch, and the rights holders who suddenly have a very concrete reason to fund them.

Shared on Bluesky by 2 AI experts