AI disclosure and watermarking

What is marked, and what is not

• Audio we generate - a read, a performance, a cloned voice, a dubbed track - carries an inaudible AudioSeal watermark and downloads with a synthetic-media tag in the file itself.
• Streams are not watermarked: a watermark is embedded over a finished file, and a stream has no end while it is playing. The response says so in its headers, so a caller can never be misled about which it has.
• Audio you sent us - transcription, cleanup, the source of a voice change - comes back as your own recording, unmarked. We do not stamp a file we did not synthesise.

The exact wording for your account, and the mark its files carry, is served at /v1/disclosure so a publishing pipeline can read it rather than copy it. /v1/verify reads the watermark back out of a clip.

What you must say when you publish

Tell your audience the voice is synthetic wherever a listener could reasonably take it for a person: advertising, customer service, narration presented as a named human, anything political and anything aimed at children. A line in the description, a spoken tag, or an on-screen label all count.

In the European Union, Article 50 of the AI Act puts that duty on you as the deployer of the system: content that is artificially generated must be disclosed, in a clear and distinguishable way, at the latest at the first interaction. We provide the machine marking and the wording; the publishing decision is yours, and our acceptable use policy makes it a condition of the licence.

Voices of real people

A cloned voice is created in the dashboard rather than over the API, because a consent record has to be signed by a person: who the speaker is, that they agreed, and when. We keep that record with the voice and delete it with the voice.

Cloning a voice you have no permission to use is a breach of the licence. Report one to [email protected] with the clip, and we will read the watermark, identify the account and act on it.

What a watermark does not prove

A mark survives normal encoding, trimming and level changes; it does not survive a determined attacker who re-records the audio through the air or filters it aggressively. A negative result therefore means "no mark found", not "not generated by Naadly". We would rather say that plainly than let the mark be read as proof it cannot be.