How text watermarking differs
Google explains that SynthID Text influences token-selection probabilities during generation. There is no image pixel grid or file metadata layer to process, and ordinary image techniques do not apply.
No text remover in this beta
SynthID Remover accepts only JPEG, PNG, and WebP images. It does not accept pasted text, documents, audio, or video, and this page does not claim that rewriting text reliably removes a watermark.
Detection has important limits
A text detector result should be interpreted within the originating product, language, passage length, and supported model. Absence of a recognized signal does not prove human authorship.
For the current primary description, visit Google DeepMind’s SynthID overview.