Google DeepMind launches SynthID Bio to watermark AI-made proteins

Google DeepMind has introduced SynthID Bio, a tool that embeds a watermark, a hidden identifying mark, into proteins designed by artificial intelligence. The company says the watermark can be detected later without changing how the protein folds or functions.

The tool extends the SynthID family, Google DeepMind’s existing line of watermarking technology, into biology. Previous SynthID tools marked AI-generated text and images. SynthID Bio applies the same basic idea to proteins: a sequence of amino acids that AI software designs from scratch, rather than a sentence or a picture.

In its announcement, titled “Introducing SynthID Bio,” Google DeepMind linked the new tool to AlphaProteo, its model for generating novel proteins for biology and health research. AlphaProteo produces protein designs that researchers can use in drug discovery and other scientific work. SynthID Bio appears to be aimed at marking proteins that come out of tools like AlphaProteo, so that a synthetic protein can later be traced back to its AI origin.

Why a watermark matters for proteins

A protein’s function depends on its exact structure, so any watermarking method has to avoid disrupting that structure. Google DeepMind says SynthID Bio is built to preserve biological function while still leaving a detectable trace that the sequence was AI-generated. The company has not published, in the material reviewed for this story, the technical details of how the watermark is embedded or how it is later read back out.

The broader goal, based on Google DeepMind’s framing, is traceability. As AI tools make it easier to design new proteins, being able to confirm that a given protein sequence originated from an AI model could matter for researchers, regulators, and biosecurity reviewers trying to track synthetic biology work. Google DeepMind positions SynthID Bio as part of that tracking effort, though the specifics of who can access the detection tool, and under what conditions, are not detailed in the announcement text available here.

Several open questions remain unanswered in the material published so far. It is not clear how resistant the watermark is to modification if someone alters a protein sequence after it is generated, what detection accuracy Google DeepMind has measured, or whether the tool will be released for outside researchers to use directly. The announcement also does not specify a release timeline beyond the introduction itself.

What to watch next is whether Google DeepMind publishes a technical paper or documentation laying out how SynthID Bio’s watermarking method works and how it performs under testing, and whether the tool is made available to the broader research community working with AlphaProteo and similar protein-design systems.