Function-preserving watermark marks AI-generated proteins

On September 30, 2026, Nature's machine learning section carried two related reports. One described a study proposing a function-preserving watermarking method for AI-generated proteins, embedding a watermark while keeping the protein's biological function intact so that its AI origin can be marked. The other described a secret watermark that labels proteins as "AI-generated" and said it offers a way to trace AI-generated proteins. Neither report named the researchers or the technique.

Later that day, at 23:03, Google DeepMind announced SynthID Bio, which it described as a proof of concept for watermarking AI-generated proteins while preserving biological function, giving synthetic biology a provenance capability. According to the announcement, the watermark can be embedded in AI-generated proteins without affecting their biological function.

At 23:54 the same day, Ars Technica reported that Google had developed a watermarking technique for AI-designed proteins. The technique embeds an identifiable marker without breaking the protein's function, making it easier to trace where a protein came from, and is intended to help with biosecurity. Ars Technica said it works with a popular AI protein design tool, but did not name that tool.

The accounts differ in places. The two Nature reports do not mention Google or DeepMind and give no name for the technique, while the DeepMind and Ars Technica items tie the work to Google and to SynthID Bio. The materials do not say whether the Nature study and SynthID Bio are the same piece of work, and they give no details on how the watermark is embedded or how reliably it can be detected. As of the last report, DeepMind still describes the technology as a proof of concept.