Contents
A SynthID check on a song tells you one narrow thing, whether Google’s invisible SynthID watermark, stamped at generation time, is present in the audio. It is not a general AI-music detector, and it says nothing about a track from a generator that never applied the mark. Read a hit as “a SynthID-adopting tool made this,” and read a miss as “no SynthID watermark was found,” never as proof that a human performed it.
What does a SynthID check on a song actually read?
SynthID embeds an inaudible pattern into generated audio. Google DeepMind describes the audio mark as inaudible to the human ear and says it “can’t be altered by common modifications like adding noise, MP3 compression, or changing the speed.” The check answers one yes-or-no question, is that pattern here, and nothing more. Gowal, Bunel, Stimberg et al. (DeepMind 2025) draw the line for the whole SynthID family: “establishing provenance is materially different from detecting AI-generated content.” A SynthID check is a probe for one watermark, not a classifier that hears AI-ness in the music.
Which music carries SynthID, and which does not?
Only audio from providers that opted in carries the mark. Google applies SynthID to its own Lyria music generator and to NotebookLM audio, and tracks generated with Lyria in the Gemini app are embedded with it. Music from Suno, Udio, and other systems that have not adopted SynthID carries no such mark, so a SynthID check on a Suno or Udio track returns “not detected.” That is the central limit for music specifically: the two names most people mean by “AI song” are exactly the ones a SynthID check cannot see. A miss is the normal result for most AI music in circulation, not evidence of a human.
What does a hit mean, and what does a miss mean?
| Result | Most likely meaning | Does not mean |
|---|---|---|
| Watermark found | A SynthID-adopting tool (Google Lyria) made it | The track is unedited or human-performed |
| Not detected | No SynthID mark is present | A human made it, or it is not AI |
A hit means a SynthID-adopting provider generated the audio, not that the track is untouched and not that Google in some general sense is responsible. A miss means only that this one watermark family was not found by this one tool.
How strong is the evidence for SynthID on audio?
Weaker than for images, and worth stating plainly. SynthID-Image has a published benchmark: an external variant reached a 99.72% true positive rate in the aggregated worst-case setting at a fixed 0.1% false positive rate (Gowal, Bunel, Stimberg et al., DeepMind 2025). SynthID-Text was documented in Nature, where Dathathri, See, Ghaisas et al. (2024) assessed “nearly 20 million” Gemini responses in a live experiment. SynthID applied to music and audio has no equivalent public technical paper, so its robustness figures are vendor claims rather than independently benchmarked ones. An audio SynthID check is therefore more provider-bound and less publicly evidenced than the image one, and the claim that it survives MP3 or a speed change is Google’s own, not a third-party result.
Where can you actually check a song?
Access is limited and fragmented. Google’s consumer route is to upload the clip to the Gemini app and, in DeepMind’s words, “ask if it’s been created or altered by Google AI.” Google’s dedicated SynthID Detector portal is not a universal public scanner; it is in early testing with journalists and media professionals, and for audio it pinpoints the specific segments where a SynthID watermark is detected. No detector outside Google reads SynthID, and SynthID cannot read anyone else’s mark, so “not detected” almost always means “not detected by this tool, for this watermark,” rather than “no watermark exists.”
A watermark is one signal, not a verdict
Because SynthID is cooperative, absent from most AI music, and, like any watermark, forgeable, a check is one input rather than a ruling. Golaszewski, Krawetz and Sherman (2026) make the general point that provenance systems provide “provenance signals, not proof of authenticity,” and watermark research has shown marks can be scrubbed and even copied onto innocent files. Jovanović, Staab and Vechev (ICML 2024) reported scrub-and-spoof attacks on state-of-the-art watermarks “for under $50” with an “average success rate of over 80%.” Treat a SynthID hit as one corroborating signal, weigh it against the track’s history and a detector that works differently, and read a miss as the absence of a cooperating mark, not a clearance. Since a miss is the normal result for most AI music, the actual detection route is is this song AI-generated.
Sources
- Gowal, Bunel, Stimberg et al. (2025). SynthID-Image: Image Watermarking at Internet Scale. Google DeepMind.
- Dathathri, See, Ghaisas et al. (2024). Scalable Watermarking for Identifying Large Language Model Outputs. Nature.
- Google DeepMind. SynthID.
- Google DeepMind. Lyria.
- Golaszewski, Krawetz, Sherman (2026). Verifying Provenance of Digital Media: Why the C2PA Specifications Fall Short.
- Jovanović, Staab, Vechev (2024). Watermark Stealing in Large Language Models. ICML 2024.