Google DeepMind wants a way to tell whether a protein was designed by a model or evolved the old-fashioned way.
The lab announced SynthID Bio on September 30, 2026, describing it as a proof of concept for watermarking AI-generated proteins. The method embeds a detectable signal into a protein's design while preserving its biological function, meaning the watermark should not stop the protein from doing its job. It extends the SynthID approach DeepMind has already applied to other AI outputs into the much higher-stakes territory of biology. At this stage, DeepMind is calling it a proof of concept, not a shipped tool.
This matters because AI-assisted protein design is moving fast, and so are the dual-use risks that come with it. A model that can design useful enzymes or therapeutics can, in principle, be pointed at more dangerous sequences, and biosecurity researchers have been asking for ways to trace synthetic biology back to its source. A watermark that survives the design process without wrecking function is a plausible piece of that traceability puzzle.
The catch is the same one every watermarking scheme runs into: it only works if the labs, tools, and screening services across the industry actually adopt it. A proof of concept from one company is a start, not a safeguard.