Google built a text-to-music AI called MusicLM, but won't release it

If you ask an AI to generate an illustration or churn out some Python code, nobody bats an eye anymore. But what happens if you type: "enchanting jazz song with a memorable saxophone solo and a solo singer"?

Until now, the results were usually a garbled mess of digital screeching. But Google just unveiled MusicLM, a new AI system capable of turning detailed text prompts into remarkably coherent, high-fidelity music across virtually any genre.

It is so good that researchers and early testers on X called it a bigger leap than ChatGPT, claiming Google has "almost solved music generation." But despite the breakthrough, Google has no plans to release it to the public anytime soon.

280,000 hours of training

AI song generation is not brand new. We have seen experiments like Riffusion, Dance Diffusion, OpenAI's Jukebox, and Google’s own AudioML. But due to technical bottlenecks and limited datasets, none could handle complex compositions or produce high-fidelity audio.

MusicLM changes that equation. Trained on a massive dataset of 280,000 hours of music, the model outputs 24 kHz audio and can interpret nuanced, multi-layered prompts, like "Berlin 90s techno with a low bass and strong kick."

Even wilder is how it handles multi-prompt workflows. MusicLM can take a hummed melody or a whistled tune and build an entire orchestral or synth arrangement around it. It can even take a sequence of descriptions—say, moving from a "meditation song next to a river" to "electronic song played in a videogame"—and stitch them into a continuous, cinematic soundtrack that shifts moods over several minutes.

It can set the "experience level" of the generated musicians, adapt to specific historic eras, or craft songs based purely on an image and a caption.

That said, it is not flawless. Some tracks carry noticeable audio compression artifacts, and its synthesized vocals are mostly incomprehensible gibberish that sounds like a blend of different voices singing fake words.

The 1% problem holding it back

So why is Google keeping MusicLM locked behind closed doors? The short answer is copyright.

In their research paper, Google’s team noted that during stress tests, about 1% of the music the system produced was directly replicated from the copyrighted songs in its training data.

In text, 1% might look like an accidental cliché. In music, a 1% direct rip of a melody or riff is an open invitation for a massive lawsuit from record labels. Google explicitly acknowledged the "risk of potential misappropriation of creative content" and opted not to release the model in its current form.

Music industry lawyers are already sharpening their knives. Eric Sunray, a legal intern at the Music Publishers Association, argued in a whitepaper that systems like MusicLM violate reproduction rights under the US Copyright Act by digesting existing works to weave coherent audio. Meanwhile, technologist Andy Baio pointed out that AI music would likely be treated as derivative work, leaving commercial usage in deeply murky legal waters.

Why it matters

For creators and game developers, MusicLM shows that custom, high-quality audio on demand is finally within reach. The immediate value here is not replacing pop stars, but wiping out the friction of stock audio: anyone needing a background loop for an indie game, video, or podcast could soon get custom music tailored down to the exact bpm and mood.

However, we believe keeping the model locked down is both a smart legal shield and a strategic vulnerability for Google. It avoids the immediate fury of the music industry—an industry notorious for aggressive litigation, as seen back in 2020 when Jay-Z’s label hit YouTube over deepfaked vocals.

At the same time, this conservative playbook leaves a massive door open for open-source rivals. Just as Stability AI capitalized on tech giants hesitating over image generation, another scrappy player will likely train an unrestricted music model without caring about that 1% duplication rate. MusicLM proves the tech works; now, the real battle shifts from computer science departments to federal courtrooms.