SynthID and Content Credentials can label AI audio. Neither can prove your song is human

On 18 February 2026 Google put Lyria 3, its music model, into the Gemini app in beta, letting anyone aged 18 or over make 30-second tracks from text or a photo. One line in that announcement matters more to working musicians than the model itself: every track made in the app carries SynthID, Google's inaudible watermark. You can also upload a file to Gemini and ask whether Google AI made it.

Meanwhile, a second system is spreading: C2PA Content Credentials, a signed record of where a file came from. Both are now pulled into the same argument, which is "is this song AI?" They answer different questions, and neither answers the one many artists will soon be asked: can you prove you made this?

What does SynthID actually tell you?

SynthID is Google's watermark for content from its own models. DeepMind's SynthID page says it covers audio made with Lyria and NotebookLM's podcast feature, and that the mark survives "common modifications like adding noise, MP3 compression, or changing the speed of the track". You can check for it in Gemini, Search and Chrome. Google's SynthID Detector portal also reads marks from partners, which the page names as OpenAI, NVIDIA and Kakao, with Apple listed as coming soon.

So a positive result is strong, specific evidence: a Google model, or a listed partner's, made or touched this file. A negative result tells you almost nothing. A model outside that group leaves no SynthID mark at all, and Google's robustness claim is limited to "common" edits, not determined attempts to remove it. Google also says Gemini combines the watermark check with "its own reasoning" when it answers, so the reply is not a pure watermark readout.

What are Content Credentials?

C2PA is an open standard whose steering committee includes Adobe, Google, Sony, TikTok, OpenAI and the BBC. Its idea is a "nutrition label" for media: a signed manifest that records who created a file, what tools and actions were used, and which ingredients went in.

Audio is properly covered. Version 2.2 of the technical specification sets out how to embed manifests in WAV and Broadcast Wave files, in MP3 and FLAC through ID3 tags, and in MP4-family files. The open-source SDK lists wav, mp3, flac and m4a as supported. Manifests can carry a "digital source type" from the IPTC vocabulary, which separates "Created using Generative AI" from "Digital creation", defined as media created by a human using non-generative tools.

The standard is honest about its limits. The C2PA explainer says provenance "cannot tell you whether the digital content is true, accurate or factual", that provenance is not always complete, and that a manifest can be stripped from a file. Its answer to stripping is "durable" credentials, which use a watermark or fingerprint lookup to find the record again.

Why can't either one prove a song is human?

Because each proves something narrower.

• No watermark is not a human signature. SynthID only exists where a participating model put it. Music from any other generator, or a human recording, looks the same to the check: nothing found.
• A credential is a signed claim, not a fact. A "digital creation" label means someone signed a statement saying so. The explainer says C2PA makes no judgement on whether provenance data is "true".
• Identity is a bolt-on. The core spec does not directly address human identity. The Creator Assertions Working Group's identity assertion, version 1.2 (ratified 15 December 2025), lets a creator link a manifest to a website or social account they have proved they control. Even so, it says the assertion should not be read as conveying "attribution or ownership" of the file.

Put plainly, these tools can help show a file came from AI. They cannot show the reverse. If someone questions your work, the strongest evidence is still the record of how you made it. Our guide to why AI detectors are not proof covers the same gap from the detection side.

How can an artist use provenance to show their own work?

You can't prove a negative, but you can build a record that is hard to fake and easy to show.

1. Keep the session, not just the master. Save DAW projects, raw takes, voice memos and MIDI with their original dates. Don't flatten or tidy them after release.
2. Export stems at milestones. Demo, arrangement, final mix. Whatever you work in, from a desktop DAW to a browser DAW such as Veena, which exports WAV, MP3, MIDI or every track as stems, archive those exports with the date in the file name.
3. Store a copy somewhere that timestamps it. Cloud storage or an email to yourself gives an independent date that you didn't write.
4. Sign masters where your tools support it. If your software can add Content Credentials, use it for the final WAV and record the actions honestly, including any AI-assisted steps.
5. Check what you shipped. Open the file in the CAI's Inspect tool, which reads WAV, MP3 and M4A. You should see a "Recorded by" line naming the signer and a list of actions. A signer not on the C2PA trust list shows as "Unrecognized". A file with no credential shows "No Content Credential", which is what happens if a platform strips it.
6. Link your identity if you can. Where a tool offers a CAWG identity assertion, tie it to your own website or a public profile.
7. Disclose AI use plainly. A provenance record that hides AI steps weakens your case. Our guide to DDEX and Spotify AI credits shows where to declare it.

Provenance pack checklist

• Original session files with creation dates
• Dated stem exports at each stage
• Raw recordings (vocals, instruments, voice memos)
• A copy in timestamped storage
• Signed master, checked in Inspect
• Notes on any AI tools used, and for what

What we think

Watermarks and Content Credentials are useful, and Google deserves credit for marking every Lyria track by default. But both are built to label machines, and the people they put under pressure are humans.

The burden is pointing the wrong way. If "no watermark" can't clear a track, artists end up proving a negative while the generators decide what gets marked. Watermarking only works if every major model takes part, and today the check is run by the companies that make the models. And as we reported, independent musicians are suing Google, alleging Lyria was trained on their music.

Process beats labels. Our analysis: the artists best placed in an authorship dispute will be those with dated sessions and stems, not those with the cleanest metadata. Content Credentials help most when they sit on top of that record, not in place of it.

Platforms are the missing link. Credentials that platforms strip on upload, or never show to listeners, protect no one. Distributors and streaming services should keep manifests intact and display them. Until they do, keep your own provenance pack.

This is general information, not legal advice. For an authorship or copyright dispute, talk to a lawyer where you live.