A group of independent musicians dragged Google into court this spring with a simple claim. The company scraped songs they posted to YouTube, they say, and used them to teach its Lyria 3 music model how to make new ones. Google's response, filed as a motion to dismiss, didn't bother denying it. The argument, more or less: even if it did, the plaintiffs already signed away the right to complain.
That distinction matters more than it sounds. Google isn't saying "we didn't." It's saying you can't prove we did, and it wouldn't be illegal anyway. Two very different sentences, stacked into one legal shrug.
What the musicians actually allege
The lawsuit, reported by The Verge in June 2026, accuses Google of feeding music uploaded to YouTube into Lyria, its model for generating audio on demand. The artists never agreed to be training fuel, the argument goes, and their work helped build a system that now competes with them.
Google's filing goes after the foundation of the case. It calls the claim an untested theory, says the plaintiffs can't show their specific songs were used, and argues that even taken at face value the complaint collapses. Why? Because everyone who posts to YouTube clicks through a Terms of Service agreement handing Google a sweeping license to use what they upload.
That license, Google notes, covers the right to reproduce content, distribute it, and prepare derivative works. Which, if you squint, sounds an awful lot like a permission slip for AI training, written years before AI training was the point.
It's the kind of clause nobody reads and everybody accepts. And now it's load-bearing.
The Lyria AI training question Google won't answer
Here's the genuinely strange part. Asked point-blank whether it uses YouTube videos to train Lyria 3, Google declined to comment.
Not a denial. Not a confirmation. Silence.
That's odd, because the company has been pretty open about doing exactly this with its other models. In April 2024, YouTube CEO Neal Mohan told Bloomberg a slice of the platform's videos might be used internally to train systems like Gemini. A company blog post later that year said the same thing in plainer terms, confirming that uploaded content gets used to sharpen products across YouTube and Google, AI tools included. Google even told CNBC, on the record, that YouTube uploads feed both Gemini and its Veo video model.
So the pattern is established. The one model Google has pointedly never named in this context is Lyria.
My read: that gap isn't an accident. Confirming the Lyria training pipeline right now, mid-litigation, would hand the plaintiffs a quote they'd love to wave in front of a judge. Saying nothing keeps a door open. Lawyers call it preserving your options. The rest of us might call it not wanting to be caught on tape.
Why "you agreed to the terms" might actually hold up
It's tempting to roll your eyes at the terms-of-service defense. We all click "I agree" without reading. Surely that can't bind us to having our creative work melted down into model weights?
Maybe not. But courts have generally treated those clickwrap agreements as enforceable, and YouTube's language is broad on purpose. "Derivative works" is a famously elastic phrase. Whether training an AI on a song counts as preparing a derivative work is one of the big unsettled questions hanging over the whole industry, and Google would clearly rather win on the contract than fight that one on the merits.
The catch: a license to reproduce and remix content so a video platform can run isn't obviously the same as a license to build a commercial product that generates competing music. Those musicians uploaded songs to be heard, not to be studied by a machine that would eventually do their job. A judge could decide the terms stretch that far. A judge could also decide they don't.
That's the whole ballgame. And it won't be settled quickly.
The bigger fight this is a piece of
Google is hardly alone. Nearly every major AI developer is fending off some version of the same accusation: it built a lucrative model on a mountain of work it never paid for. Music is just a particularly raw nerve. The people affected tend to be individuals rather than corporations, and the output, a freshly generated song, lands uncomfortably close to the input.
What makes the YouTube angle distinct is the scale, and the captive supply. Google doesn't have to scrape the open web for music. It owns the largest video platform on earth, and millions of creators hand it material voluntarily, under terms Google itself wrote. If that arrangement holds up in court, it's a structural advantage no rival can copy. Spotify can't train on YouTube's library. Google can.
Which is also why the artists' frustration is easy to understand. They weren't tricked, exactly. They just never imagined the deal they accepted to reach an audience would later be cited as consent to compete with them.
What to watch
The immediate question is whether the motion to dismiss works. If the judge buys the terms-of-service argument and tosses the case early, it sets a powerful precedent for every platform sitting on a pile of user uploads. Expect Meta, TikTok, and anyone else with a content firehose to take notes.
If the case survives, things get more interesting. Discovery could force Google to say on the record what it won't tell reporters now: which models were trained on what. And that's the moment the company has been quietly working to avoid.
Worth keeping an eye on, too, is a separate thread: regulators. The same week this story surfaced, Google was also expanding the kinds of user data it retains for AI training across other products, from Lens photos to Translate audio. The direction of travel is clear enough. More collection, broader terms, less friction. Whether the courts or lawmakers push back is the open question, and right now nobody can say how it lands.
For the musicians, the stakes are simple. If a broad license buried in a sign-up flow really does cover this, then uploading anything to a major platform increasingly means feeding the machine that may replace you. Plenty of creators would have made a different call if someone had spelled that out. Now a court gets to decide whether it matters that no one did.