DeepMind Puts a Hidden Watermark on AI-Designed Proteins

Google DeepMind has released SynthID Bio. It is a method that embeds a hidden watermark directly into the amino acid sequence and the 3D structure of an AI-designed protein. The team tested the watermark on three proteins, including a binding domain from the SARS-CoV-2 spike protein. The watermark was detected almost every time. Binding strength and the diversity of natural sequences stayed the same in watermarked proteins compared to regular ones. The method has already been tested on live bacteriophages, too, and they stayed fully functional even with the watermark inside.

This has nothing to do with writing code, but the logic matches watermarks in text and images. The more things AI creates, the more it matters to know what it made and what it did not. In biology the cost of a mistake is higher, so the method is stricter, too. The watermark survives actual synthesis of the protein in a lab, not just inside a file on a computer. For now this is code and data for researchers, not a finished service, so you cannot apply it to your own work right away.