Detection works in some cases, and the tools are improving. Detection, watermarking and provenance answer different questions. A result from any of them is not a legal verdict.
These technologies can strengthen transparency around AI-generated audio, but they operate differently and have different evidential value.
Analyses characteristics of the audio itself to estimate whether it may have been generated by a known AI system.
Embeds an identifiable signal at generation time. A detected watermark can support provenance; its absence does not prove human authorship.
Cryptographically records where media came from and how it has been modified, rather than attempting to infer origin from the sound alone.
An AI-detection score, on its own, does not establish infringement, ownership, whether training was lawful, or whether a production has the rights for commercial use. Detection and provenance can raise those questions months or years after the work is in circulation. Clearance is a separate check from whether a detector flagged the file at release.
Useful for triage and investigation. Their accuracy depends on the models covered, the audio transformation applied and the evaluation methodology.
IRCAM Amplify offers an AI Music Detector designed to analyse audio and return an indicative detection result. Act in Synch treats any such result as an investigative signal, not a legal conclusion.
Official documentation ↗Deezer says its detector identifies fully AI-generated music with 99.8% accuracy and has made a public playlist-checking tool available alongside a professional solution.
Deezer newsroom ↗In May 2026, the European Commission published a dedicated technical study on marking and detecting AI-generated audio in the context of Article 50 of the AI Act.
European Commission ↗A watermark is added by the generator. A detector tries to infer the origin afterwards.
Google DeepMind says SynthID embeds an imperceptible watermark into audio generated or published through Lyria. The watermark is designed to remain detectable through common transformations such as MP3 compression, added noise or speed changes.
Google DeepMind ↗A positive SynthID result can indicate Google AI provenance. A negative result does not establish that the music was made by a human or that another generative model was not used.
C2PA Content Credentials provide an open standard for recording and validating the origin and history of digital media, including audio. For generative AI, credentials can identify output as trained-algorithmic media and can carry information about models, inputs and processing history.
This records signed information about where the file came from and how it was changed. It can be checked when the credentials are present and intact.
C2PA AI/ML guidance ↗A track may be identifiable as AI-generated without telling you whether the training data was licensed. Conversely, a model can document a consent-based training approach without that proving anything about a specific output’s exclusivity or clearance.
Founded by Ed Newton-Rex, Fairly Trained certifies qualifying generative-AI models whose certified training datasets meet defined licensing, consent, open-licence, public-domain or ownership criteria.
Certification criteria ↗A certification can be relevant evidence about a model developer’s data-sourcing approach.
It does not by itself establish output ownership, exclusivity, client transfer, media rights or provider indemnity.
A detector cannot determine infringement by itself.
Ownership depends on law, contracts and the circumstances of creation.
Detection of AI generation does not reveal the legal basis for training.
That depends on the provider terms and the production contract.
Indemnity and liability allocation are contractual questions.