Google DeepMind makes its AI text watermarking available as open source

SynthID introduces additional information at generation time by changing the probability of tokens being generated, explains Kohli.

To detect the watermark and determine whether text was generated by an AI tool, SynthID compares the expected probability values ​​for words in watermarked and unwatermarked text.

Google DeepMind found that using the SynthID watermark does not affect the quality, accuracy, creativity, or speed of the generated text. This conclusion was drawn from an extensive live experiment on SynthID's performance after the watermark was deployed in its Gemini products and used by millions of people. Gemini allows users to rate the quality of the AI ​​model's answers with a thumbs up or a thumbs down.

Kohli and his team analyzed the results for around 20 million chatbot responses with and without watermarks. They found that users noticed no difference in quality and usefulness between the two. The results of this experiment are detailed in an article published in Nature Today. Currently, SynthID for text only works with content generated by Google models. However, there is hope that the open source solution will expand the range of tools it is compatible with.

SynthID has other limitations. The watermark was resistant to some manipulations, such as: Such as cropping text and light editing or rewriting, but it was less reliable when rewriting AI-generated text or translating it from one language to another. It is also less reliable in responding to prompts that ask for factual information, such as the capital of France. This is because there are fewer ways to adjust the probability of the next possible word in a sentence without changing the facts.