Transcription

• Speaker #0

  Next time, the ones who ask the question. That was the promise we left you with last week. This week, the question changed hands. Every single day, a French music platform receives around 90,000 tracks, made entirely by machines. On some days, that's more than half of everything that comes in. This isn't a forecast. These are the platform's own figures, and almost nobody listens to them. Between one and three streams out of a hundred. That same week, a man was sentenced to 18 months in prison for getting his artificial songs played by thousands of fake listeners. So, who are these songs actually made for? And who does the sorting? Welcome to the week of October 12th, 2026, Synth AI Cast, the AI that listens for you. Episode 29, the 7th of Season 2. This week, a flood that nobody listens to, a fraud that fed on it, and sorting tools that are finally landing in your hands. Last week, we talked about the invisible mark that... that ChatGPT is about to put on the text it writes in Europe, a mark you had no way to read. This week, for sound and images, that changes. First signal, the flood. On Tuesday, Deezer, the music streaming service born in Paris, made a simple announcement. By the end of the year, all around the world. A switch will let you clear out of your searches. The tracks, the albums, and the artists. Generated entirely by AI. Whenever you choose. Why a switch? Because of the numbers Deezer publishes itself. In January 2025, the platform was receiving about 10,000 of these tracks a day. Today, it's close to 90,000. Nine times more. in less than two years. On peak days, these tracks make up more than half of all deliveries. In other words, on some days, one track in two was never played by a musician. Since January 2025, more than 42 million of them have been pulled from recommendations. And yet, they account for only one to three streams in a hundred. half of what arrives, for a few percent of what people actually hear. Both numbers are true at the same time, and the gap between them is the real story. Which leaves an awkward question. Can we, the listeners, even tell the difference? Last year, Deezer commissioned a study from Ipsos. 9,000 people, across 8 countries, in a blind test. They heard three tracks, two made by AI, one made by musicians. 97% couldn't say which one was human. In the same study, almost half of streaming users said they wanted to be able to filter that music out. The switch, in fact, only targets tracks that are fully generated, not artists who use AI as one tool among others. Second signal. Why make music nobody listens to? On Tuesday, October 6th, in New York, a man from North Carolina was sentenced to 18 months in prison. He had pleaded guilty back in March. Federal prosecutors call it the first criminal case of its kind. The scheme, as the Justice Department describes it, comes in three moves. First, Hundreds of thousands of songs generated with AI. Then, thousands of fake accounts, up to 10,000 at a time. Finally, software that plays those songs on a loop. Billions of streams. And more than 8 million dollars in royalties collected. Those royalties come out of a shared pot. Platforms split it among artists, in proportion to their streams. So... every fake stream takes a little from a real musician. One example from the case file, with its fine print. In April 2023, on YouTube Music, and counting only family plan accounts, his fake accounts played his tracks nearly 81 million times, almost nine times more than Taylor Swift's entire catalog, on those same family plans. On Wednesday, on The Verge cast, The Verge's flagship show, Wired reporter Kate Nibbs walked through her investigation. This man had dreamed of a music career for years. According to his former business partner, he was already faking his streams long before today's tools. A, I mostly let him do it at scale. The platforms did catch on in the end. Some faster than others, according to her reporting. And she made a point that's easy to forget. Streaming pays so little that $8 million is a fortune in the music business. Which brings us back to our opening question. A good share of these tracks aren't made for ears. They're made for counters. According to Deezer, in 2025, up to 85% of streams on fully generated music were fraudulent, compared with about 8% across its whole catalogue. Third signal. The ones who ask the question, from now on. On Wednesday, October 7th, Google opened its detector to everyone. SynthID detector. Free, in English. You give it an image, a video, or a sound, and it tells you whether it carries the mark of Google's AI, or one of its partners, OpenAI among them. An earlier version had been around since last year. reserved for media professionals. The question has moved from journalists to everyone. Google says it has already marked the equivalent of 240,000 years of audio, but it only recognizes those marks, nobody else's. And above all, not text. For text, the detectors from Anthropic and OpenAI are still restricted, by application only. to a few organizations. The same day as Deezer, Meta took a step in the same direction. On Facebook, a new feature lets you say, in your own words, what you want to see more of, or less of. Language models interpret the request, and your feed changes within the same session. Same idea. You talk to the algorithm, instead of it guessing about you. What Meta... doesn't say, though, is what it does with what you type. Let me take you to Paris in the spring of 1863. The Jury of the Salon, France's official painting exhibition, turns down a large share of the canvases. The artists protest so loudly that the emperor steps in. The rejected works will be shown separately, and the public will judge for itself. It becomes known as the Salon des Refusés, the Exhibition of the Rejected. Manet shows a painting there that we now know, Le Déjeuner sur l'Herbe, the Luncheon on the Grass. Public bursts out laughing. The sorting hadn't disappeared. It had changed hands, and nothing guaranteed those hands were any steadier than the jury's. In 1863, the jury decided for everyone. In 2026, the switch decides for you, if you turn it on. Last week, we asked who would have the right to ask the question. This week, the answer is everyone, or almost. For sound and images, you can now ask, was this made by a machine? For music, you'll soon be able to brush it aside with a single tap. For text, you still can't. 90,000 a day. That number doesn't say AI is replacing musicians. It says AI produces faster than we can listen. And when we can no longer listen to everything. Whoever does the sorting decides what exists. What to watch for next. The actual launch date of Deezer's Switch. What Spotify and Apple Music decide to do. and the first text detector open to everyone. Three things to try this week. Five minutes. On Deezer, check whether an album carries the label that flags content generated by artificial intelligence. Thirty minutes. Run a recent photo or sound clip through Google's detector. One hour. Listen to the Vergecast episode on this fraud. It runs just under 40 minutes. This week, the sorting started coming back to you. When machines produce faster than we can listen, who chooses what reaches you? Next time, the ones who choose for you.