NewsResearch
Google DeepMind's SynthID Bio watermarks AI-designed proteins so synthesis screeners can tell where an order came from
Google DeepMind published SynthID Bio in Nature on October 1. The method hides a signature in the amino-acid sequence and the predicted 3D coordinates of proteins designed by an AI model, so that DNA synthesis providers screening orders can check whether a design came from a trusted, safety-vetted model. The watermark is built into the model's weights, so every output carries it. In wet-lab tests on three protein binders, watermarked designs matched unwatermarked ones on hit rate, binding affinity and diversity. DeepMind says resistance to deliberate tampering is unresolved and recommends layering the watermark with provenance metadata, repositories and customer vetting.
Why it matters
Provenance for AI output is arriving where the stakes are highest. The same idea, a mark that says which model produced something, is what code and document pipelines will need.