Suno made AI music a firehose. Now, facing lawsuits, it wants to watermark the flood

The plans landed in a blog post on Thursday from Suno’s co-founder and chief executive, Mikey Shulman. The company will add audio watermarking and fingerprinting. It will roll out transparency tools so other platforms can spot its tracks. And it will bring in a new downloads policy to curb mass distribution.

The Verge first reported the move.

Shulman framed it around a set of principles. Chief among them: “great music is made by people.” The watermarks, he wrote, are “durable and resistant to tampering.”

They will not judge “whether a song is good, meaningful, or sufficiently human,” he added. Whether to label a track as AI should be the call of artists and platforms, not Suno.

What is actually changing

Every track Suno makes would carry an invisible signature in the waveform. Partner platforms could then flag or block it. Suno did not say whether it will build its own system or license one. Google’s SynthID is the obvious option, and OpenAI and Nvidia already use it. Suno also named the lyrics firm Musixmatch to screen outputs for copyrighted material.

The downloads change targets abuse most directly. Suno will limit mass distribution to streaming services. It says the shift will spare “the vast majority” of users. But it left out the detail that matters most: how many downloads, and at what price. The watermarking has no firm date beyond “the coming weeks.”

Why now

The timing is not subtle. Universal and Sony are suing Suno over its training data. Last week a German court ruled it broke the country’s licensing rules. A 2025 hack exposed something worse.

That hack showed Suno had scraped YouTube, Deezer, and Genius to train its models. It leaked data on a reported 55 million users. A class action followed.

So the move reads less like an ethical awakening than a bid for legitimacy. Ars Technica called it Suno “going legit.” The company raised $400m in June at a $5.4bn valuation. Now it wants to look like a tool for musicians, not a slop machine for fraud.

The fraud is real. One man recently pleaded guilty to using AI songs and bots to pull in more than $8m in fake streaming royalties.

The catch

Watermarks are not a cure. Break the signature, as Ars notes, and every track made until then can lose its label. Rival generators do not watermark at all. Open models let anyone make unlabelled music on their own machine.

The move also aligns neatly with the EU’s new AI labelling rules. Those make disclosure less a choice than a coming requirement. Streaming services from Spotify to Deezer, and platforms like Snapchat, have spent months trying to hold back the AI tide their own openness let in.

For its first years, generative AI measured success by how much it could make. Suno’s week points to a new test: how much of that flood a company can hold back. Ideally, before a court makes it.