Deepfake Laws by State: Sony Fights 260,000 AI Fake Songs

Somewhere right now, a fan is streaming what sounds like a brand-new song from an artist they love. The voice is right. The style is right. But the artist never sang a single note of it. According to Sony Music, the company has now asked streaming platforms to pull more than two hundred sixty thousand of these A.I. fakes.

That number was roughly half as big just six months

That number was roughly half as big just six months earlier. And if you've ever pressed play on a song, a voice memo, or a video clip, this story is about you. It's not really about celebrities. It's about whether anyone, a person or a machine, can still tell a real voice from a copy. According to Music Business Worldwide, the fakes imitate stars like Harry Styles, Adele, and Michael Jackson. Sony told the Financial Times that keeping up is becoming an uphill struggle. That's a major record label with deep pockets, admitting it's falling behind. So if Sony can't keep pace, who can?

Start with how fast this grew. According to Music Business Worldwide, Sony's count sat at about ten thousand in late twenty twenty-three. By March of last year, it hit seventy-five thousand. By this March, it had passed a hundred thirty thousand. Then it nearly doubled again by September. That works out to more than twenty thousand new fakes every month. Twenty thousand a month. That's roughly one new fake song every two minutes, all day, every day.

Why does that matter beyond the music business? According to Billboard, industry estimates put streaming fraud at more than two billion dollars a year. Fake tracks pull money away from the people who actually made the music. For the rest of us, it means the playlist you trust may already include voices that were never there.

The obvious answer is better detection software. Build a tool that flags the fakes automatically. The research says that's harder than it sounds. A recent study posted to the research site ar-Xiv tested top audio deepfake detectors on real-world recordings. In the lab, some of these tools catch fakes about ninety-nine times out of a hundred. Out in the real world, their accuracy dropped by about half. Some failed more often than they succeeded. In other words, a tool that aces the test can do worse than a coin flip on the job.

Researchers have a name for this

Researchers have a name for this. They call it domain amnesia. The detector learns to spot fakes made by one generation of A.I. tools. Then newer voice generators arrive, and the detector essentially forgets what a fake looks like. A tool trained on last year's fakes won't reliably catch this year's. A review in the journal Sensors adds another wrinkle. Every voice-cloning method leaves different traces behind. So there's no single fingerprint a detector can hunt for.

For investigators and lawyers, that's a serious problem. A voice recording offered as evidence can't be cleared by one automated scan. It needs a trained human checking it, plus tools that get updated as the fakes evolve. For everyone else, it means a phone call that sounds exactly like your kid might not be your kid.

And the law? It's mostly a patchwork. There's no federal law in the U.S. that stops someone from cloning an artist's voice. Protections depend on which state you're in. That leaves platforms reacting after fakes go up, instead of stopping them first.

Platforms do push back on the numbers. They point out that Sony counts removal requests, not confirmed removals. Sony hasn't said how many fakes actually came down, or how fast. Some partial fixes exist too. Spotify has artist protections. Australia's chart body, ARIA, has eligibility rules that keep certain tracks off its charts. But every one of those fixes still depends on spotting the fake in the first place.