How to tell whether a sound was created by an AI

What the exported file says

When a film exported with Cutscene plays a voice or a sound created by an AI (a voice-over or a sound effect created with ElevenLabs, or a sound the person said was created by an AI), the MP4 file says so in its metadata. An XMP packet, placed at the start of the file, holds:

• the IPTC value “Digital Source Type” http://cv.iptc.org/newscodes/digitalsourcetype/compositeSynthetic, which means: a mix of several elements, at least one of which is created by generative AI;
• an English sentence that says which: a voice-over, sound effects, or both (for example “This film contains a voice-over generated by AI.”), and nothing else.

A film in which no sound created by an AI is heard does not carry this mark. Nor does a film exported without sound (Firefox exports without sound): nothing is heard in it.

A project for After Effects or DaVinci Resolve

A project for After Effects or DaVinci Resolve is not a finished film: it is meant to be edited, and a file made from it no longer carries our mark. When a sound created by an AI is heard in the film, the project says so three times:

• at the top of its README.txt file, in English and in French;
• in the sound file, sound/mix.wav: the English sentence of the mark, in its comment (the INFO field “ICMT”);
• with a marker at the start of the sound, “AI-generated sound” or “Son généré par IA” depending on the film's language, followed by the sentence that says what was created by an AI: on the sound layer of the main composition in After Effects, on the sound clip of the timeline in DaVinci Resolve.

Its reference film (reference/film.mp4), when it has sound, carries the XMP mark described above. The film the person makes from the project does not carry it: when a voice or a sound created by an AI could pass for authentic, it is for them to say so in that film (Terms, article 13a).

How to read it

With ExifTool, a free tool that reads the metadata of files (exiftool.org), in a terminal:

If the film carries the mark, ExifTool answers “Digital Source Type : http://cv.iptc.org/newscodes/digitalsourcetype/compositeSynthetic”. If it answers nothing, the file does not carry it.

ElevenLabs' watermark

ElevenLabs says it marks the sounds it creates with an inaudible watermark. On its site, a tool you can use without an account estimates whether a voice was created by ElevenLabs: AI Speech Classifier. ElevenLabs says this estimate holds for a file that was not modified and may miss modified audio, which a mixed film is; its Audio Detector, which reads the watermark, requires an account. We have not yet checked that this watermark holds after the film is mixed and exported.

In the editor

A created voice is marked “Voice created by an AI” there, and a created sound effect “Sound created by an AI”.

The limits

• The mark is not signed: a tool can remove it, or write it into another file.
• A file copied by editing software, or re-encoded by a social network or a video platform, usually loses it. Its absence therefore does not prove that no AI was used.
• A sound brought by the person is marked only if they said it was created by an AI.
• When a voice or a sound created by an AI could pass for authentic (a real person, object, place or event), the file's mark is not enough: the film must say so clearly to whoever watches it (Terms, article 13a).