SynthID text watermark: a Gemini script in a Sume avatar video

Google DeepMind lists text from the Gemini app and web experience among the content SynthID watermarks (DeepMind, read 2026-10-02). If you paste such a script into a Sume avatar video, the finished file is a new render of speech and picture. Neither page says the text mark carries into that video, so do not assume it does.

What Google says SynthID covers

The page lists several media types under one name, which is easy to read as one mark that follows content everywhere. It is a family of marks, one per modality.

What happens to a script in an avatar render

The Sume docs describe avatar video as script-driven: you send a script or ordered video_inputs, and Sume returns a talking video. The words are turned into speech and frames. A statistical mark in the wording of text is a property of that text, and the docs do not describe it being read, preserved or re-applied.

The same goes for captions. If you burn the script as text on screen, the characters are pixels, not the original string.

What you can say honestly

• "The script was drafted with a Gemini model" is a statement about your process and is yours to make or omit.
• "The video carries a SynthID text mark" is not supported by anything on either page.
• If a platform or law asks about AI-written text, the public-interest text rules are separate from the video label; see the related post on AI-written scripts.
• Keep the original script file with the job id so you can show what was written and what was rendered.

Honest limit

Google says the detector is limited to its own tools and that access is staged. A missing mark in a Sume video does not show the script was human-written, and a present one in the script would not label the video.