Audio watermarking works by encoding a signal into a sound file at the point of generation, a signal designed to survive transcoding and repackaging so a detector on the receiving end can read it regardless of where the file traveled. That output-layer mechanism is exactly what streaming music's AI-content problem has been missing, and Suno, one of the most prolific generators of AI music on services like Spotify, says it will now build one in. CEO and co-founder Mikey Shulman announced the plan in a company blog post, citing "emerging industry standards" for AI content labeling as the reason.
What the commitment actually covers
Suno says watermarks will be applied to all audio outputs from its models. Platforms that receive those files get a machine-readable signal with two choices: display a label identifying the track as AI-generated, or block the content entirely. Shulman's post frames the move as both a technical implementation and a policy update, though it does not spell out what the policy changes cover beyond the watermarking commitment itself. The enforcement call sits with the platform, not with Suno.
Shulman did not name the watermarking technology the company will use and did not say whether Suno plans to build the capability in-house or license an existing system.
Google's SynthID as the reference point
Google recently opened SynthID, its audio watermarking technology, to outside licensees, making it the most visible off-the-shelf option in the space. Google says SynthID has been applied to audio equivalent to 60,000 years of listening time generated by its Gemini models, along with more than 100 billion images and videos. Shulman's post does not confirm whether Suno is evaluating SynthID or developing its own alternative.
The durability question
The core constraint in AI audio detection has always been that a streaming platform receives a finished file with no provenance metadata it can trust. Embedding a watermark inside the audio itself solves the transport problem, but the strength of the signal depends on the implementation. Some watermarking schemes degrade or can be stripped with basic audio processing, and Shulman's post does not address resilience against removal.
Until Suno specifies its technical approach, the durability of what it ships stays open. Google's deployment record of 60,000 years of labeled audio gives SynthID a scale reference that an internal Suno build would take time to approach.