DeepMind extends SynthID watermarking to AI-generated text and video

Google DeepMind extended SynthID, a watermarking system it had previously applied to AI-generated images, to text produced by the Gemini app and web experience and to video from its newly announced Veo generator. The text method works by adjusting the probability scores a language model assigns to candidate next tokens during generation, nudging it toward a pattern that a separate detector can later recognise statistically, without a human reader noticing any difference in the output.

DeepMind said the technique did not measurably affect the quality, accuracy or speed of generated text in testing, but it was explicit about the method’s limits: detection is less reliable on short responses, on text answering narrow factual questions, and on content that has been heavily rewritten or translated after generation. The company described SynthID as one layer of a broader provenance strategy rather than a definitive fix for identifying AI-generated content — a caveat that ran through watermarking research generally, since any watermark that survives paraphrasing without being trivially strippable remained an open problem.

The announcement came as concern about AI-generated misinformation ahead of elections in multiple countries pushed labs toward provenance tooling, and set up SynthID’s later move toward broader availability: DeepMind said it would publish further research and make the text method available to outside developers later in the year.