Provenance

SynthID and C2PA Are Not Detection

Watermarks tell you when a generator was honest. They tell you nothing when it was not, which is the case that matters.

Watermarking and content provenance are genuinely good ideas and they are advancing fast. They are also frequently presented as a substitute for detection, and they cannot be one.

The asymmetry

A watermark proves a generator chose to be honest. It proves nothing when the generator was not, and the dishonest case is the only one anybody is worried about.

What each thing actually is

SynthID

Google's watermark, embedded in audio generated through its own music models. Robust to several common modifications by the vendor's account. No true-positive or false-positive rate for the audio variant has ever been published, and detector access is restricted rather than open.

C2PA and Content Credentials

A signed manifest travelling with the file, describing how it was made. The specification covers MP4 and M4A, WAV, MP3 and Ogg. FLAC and AIFF are not covered. Real production adoption in audio is currently limited to a couple of AI vendors signing their own output. No streaming service, DAW or major label is embedding manifests on music.

AudioSeal

Meta's open, MIT-licensed audio watermark, now registered as a C2PA soft binding. Its own authors report near-perfect robustness to MP3, AAC and resampling, and a substantial drop under high-pass filtering. That the authors publish the weakness is precisely what makes it credible.

The three holes

Coverage. The largest generators have made no shipped provenance commitment. An announcement is not an implementation.
Survivability. Metadata is stripped routinely by ordinary upload pipelines, often without anyone intending it.
Adversaries. Anyone deliberately passing off generated music as human simply avoids the tools that mark it, or removes the mark.

Where regulation lands

EU AI Act Article 50(2) has required since 2 August 2026 that providers mark synthetic audio in a machine-readable, detectable format, with penalties up to 15 million euro or 3% of worldwide turnover. The accompanying Code of Practice states plainly that no single marking technique suffices, and requires at least two layers.

Notably, it says fingerprinting approaches may be preferable for audio, while stating that fingerprinting alone is not sufficient either. The regulation's own position is that this is a layered problem.

EU AI Act Article 50; EU Code of Practice on Transparency of AI-Generated Content, 10 Jun 2026

How we treat it

Provenance is read, reported, and kept in its own column, never folded into the acoustic score. A watermark found is a strong positive signal. A watermark absent is not evidence of anything, and a system that treats it as evidence will quietly accuse a lot of innocent people.

Common questions

No, and this is the most dangerous misreading in the whole area. Absence of a watermark tells you nothing. It may mean human origin, a generator that does not watermark, a stripped file, or a deliberate removal.

Partially. The specification supports several audio containers but not all of them, and there is little production deployment in music. Two AI vendors sign their own output. No major DSP, DAW or label is embedding manifests on audio.

When the tool opens

Three checks a day, free, no account.

We will write once when the benchmark is published and once when the detector opens. Nothing else.

Keep reading

The basics