🎧 Prefer to listen?

The Billboard Hot 100 may have its first AI-generated song — and the thing that exposed it wasn’t a journalist or a fan forum. It was software. Rapper Fenix Flexin’s track “Rubberz” had been under suspicion for weeks; musician Medasin named Treblo as the generator. Then Treblo itself shipped an open-source AI Music Classifier, ran the song through it, and announced the result: “very likely Treblo,” high confidence. If it holds up, the CEO is treating it as a milestone — the first known AI-generated song to reach the Hot 100, and he’s proud of it.

Detection tools work by measuring signals, not magic — underneath they’re the same plumbing of data-in, score-out that APIs explained like you’re 5 demystifies. But the arrival of a public, open-source detector with a claimed false-positive rate under 1 in 10,000 changes the math for anyone who creates content with AI in the loop. I looked at what this detector actually can and can’t do, ran its logic against the story, and came out with a playbook every solo creator should read before their next release.

What I actually did

First, the facts of the case, because the messy part is instructive. A viral track hit the Hot 100. Listeners flagged it as AI. A well-known producer publicly named the specific tool. The tool’s maker then shipped a classifier that confirmed it. The artist denied everything and posted videos of what he claims are real project files — which a producer pointed out could have been staged with an AI stem separator. Notice the sequence: denial didn’t work, because a tool existed to check the denial.

So the real question for creators isn’t “did he do it” — it’s what happens now that checking is free. I mapped out what the classifier can and can’t do:

  • It only detects Treblo. This is the part most people skimmed. The tool flags songs made with Treblo’s model specifically — not Suno, not Udio, not whatever ships next month. Detection is vendor-by-vendor, like antivirus in 2005.
  • It has a false-positive rate, not a zero rate. One in 10,000 sounds bulletproof until you’re the 1. Accusations travel at the speed of a screenshot; retractions don’t.
  • It came from the vendor itself. Treblo has every incentive to make its detections accurate — and, let’s be honest, to bask in the publicity of being the tool behind a Hot 100 first.

The uncomfortable takeaway: detection is arriving as a patchwork of one-vendor tools, and each new detector widens the gap between “could get caught” and “did get caught.”

Do this yourself (the 3-question disclosure test)

If you use AI anywhere in your creative pipeline — music, images, writing, video — run everything you publish through these three questions:

  1. Would this embarrass you if detected? There’s a world of difference between “I used AI stems and played the guitar over them” and “I generated the whole thing and claimed session footage.” The first is a workflow; the second is a lie. Fenix’s actual crisis isn’t the AI — it’s the denial plus the allegedly staged evidence. If you’d be comfortable being the headline, you’re fine. If not, change the workflow or the story.
  2. Could a tool name it? Assume every major generator will eventually ship a detector, and that third-party classifiers will cover the rest. Anything you publish stays published. If you’re not willing to have a detector flag your work in a year, don’t publish it under a false story now — this is the same fake-identity dynamic I covered in AI agents faking identities: polished surfaces hide provenance, until someone checks.
  3. Does disclosure cost you less than discovery? This is just math. Disclosing “produced with AI assistance, human-written vocals, AI-generated stems” costs you some listeners upfront. Discovery costs you all of them, plus the trust of everyone who defended you. For businesses, the asymmetry is worse — your brand runs on accumulated trust, not single posts.

Then, if you want a concrete starting point: put a one-line disclosure in your descriptions — front-loaded, matter-of-fact, no apology. “This beat was generated with [tool]; vocals and lyrics are mine” reads as confidence. A denial that later collapses reads as the end of your credibility.

What failed

  • “Post the project files” is no longer proof of anything. Clips of a session can be manufactured from AI-separated stems, as Medasin alleged here. Proof-of-work for audio is arriving (provenance standards, watermark detection) but it’s not standard yet, so videos prove nothing in either direction. Same trap as AI models hiding their own mistakes — the evidence you’re shown isn’t the whole picture.
  • The community is a better detector than the tools — for now. Listeners flagged “Rubberz” before any classifier existed. Ears notice the tell-tales: over-smooth vocals, formulaic structures, that uncanny polish. But this advantage decays fast; next year’s models will close the gap. That’s exactly the dynamic behind AI content flooding the internet — detection chases generation and always will.
  • One vendor’s detector is not an ecosystem. I checked whether any equivalent exists for the other major generators — coverage is patchy and mostly unofficial. The gap between “I used Tool X” and “no detector for Tool X” is where the next scandal is already brewing.

Who should skip this

Skip the disclosure anxiety entirely if you don’t publish creative work — a spreadsheet and a support agent don’t need provenance labels. Skip it too if you already disclose and just wanted the drama explained: the playbook above is mostly confirmation of what you’re doing. But if you’re sitting on AI-assisted work you’ve been describing as fully human — or planning to — this story is your warning shot. The tools that end careers aren’t the regulators or the journalists; they’re free, open-source, and made by the company whose model you used.

The bottom line

The first AI-generated song on the Hot 100 wasn’t exposed by an investigation — it was exposed by a detector its own generator shipped, and the artist’s denial made it worse than the tool ever could. Detection is coming for every medium, one vendor at a time, and the only reliable strategy is to be on the right side of the question before it’s asked. Run the three-question disclosure test on everything you’re about to publish. In a world where checking is free, the truth is the cheapest position you can hold.

New to AI tools entirely? Start here.