SynthID: Google makes its watermark available for AI content

In a short article on X, Google DeepMind announces its intention to make SynthID available as open source. This is a kind of watermark for AI content – and can be used to mark text, images, videos and audio. SynthID is free of charge for developers and companies. In addition to creating watermarks, it should also be able to help discover AI content. However, it is still in beta. How quickly a text or image is generated and the quality of the output should not be affected by SynthID.

To mark texts, SynthID uses the functionality of Large Language Models (LLMs). LLM responses are based on probability values that are assigned to each token (a letter, word or phrase). This probability value is adjusted and made visible by the machine. This creates a pattern. According to Google DeepMind, SynthID can be applied to as few as three sentences, but the robustness and accuracy increase with the length of a text, i.e. with more tokens.

Google is particularly interested in recognizing AI content. Not least because the internet has also been flooded with this content since the widespread availability of generative AI. It is now even harder for Google to find relevant content and display it in searches. AI and malicious actors have mastered SEO, so they know how to rank well on Google, even though the content may not be helpful to searchers. Google therefore recently ran a core update with the name "Helpful Content Update". This was primarily intended to detect AI spam. According to Google, this was a particularly complex update, although not much was revealed about the process. There was also no guidance for website operators, except that they could consider for themselves what constitutes helpful content.

However, the update also includes a new spam policy that prohibits automatically generated content without added value or original content. This also includes unverified translations and rewritten texts.

Pixel watermark invisible to the eye

Images and videos are marked with a watermark that is embedded in the pixels. This creates a pattern that is invisible to the human eye. According to Google, the watermark is quite robust, which means that it cannot be removed by a simple screenshot, for example, like metadata. The note also cannot be cut off if the image section is reduced in size.

SynthID uses the same procedure for videos. The pixels of each image are changed.

SynthID is available to people via Vertex AI, who use Google's image generators Imagen 2 and 3, as well as for ImageFX and VEO, the video AI. Only a few creators can use these. SynthID can also be used to check images for their origin. If they were created with Google's tools, this should be recognizable. This can even be done via the "About this image" query in Google Search or directly in Chrome.

(emw)