Google’s AI Image Watermark Is Edit-Resistant, Imperceptible to the Human Eye

Google experimentally launched a tool today that promises to make it much easier to tell if an image has been AI-generated. Called SynthID, it can stamp images with special watermarks that are imperceptible to human observers and supposedly resistant to most kinds of editing. The tool can also read if an image has already been stamped, and is available on Google’s text-to-image model Imagen for Vertex AI customers. 

SynthID employs two deep learning models, one to embed watermarks into AI-generated images and one to identify them. The watermarks it adds are different from traditional ones in that human beings cannot visually perceive them—the watermarks are embedded into individual pixels.

Google says the watermark remains detectable by SynthID, even if a user crops the image, adds a filter, changes the color scheme, or alters the file format. 

SynthID is limited by its platform: right now, the tool is only usable with images generated by Imagen. The Google DeepMind team sees its applicability “for use across other AI models,” writing that the group is “excited about the potential of integrating it into more Google products and making it available to third parties in the near future.”

SynthID’s watermark technology could address problems with lifelike AI, such as deepfakes, by providing an objective means of differentiating between AI content and human content. SynthID’s focus on individual pixelation means that it isn’t limited by a perceivable watermark that users can easily edit out.

Google DeepMind CEO Demis Hassabis reminded The Verge that SynthID is in beta testing and that it’s “not a silver bullet to the deepfake problem.” 

Hassabis said that once the tech is solid, “the question is scaling it up, sharing it with other partners that want it, and then scaling up the consumer solution—and then having that debate with civil society about where we want to take this.”