Google Now Hides Invisible Watermarks in AI-Designed Proteins

Google DeepMind announced a new tool called SynthID Bio that plants an invisible signature inside proteins designed by artificial intelligence. Proteins are the tiny machines that run your body — and AI can now invent brand-new ones. A watermark here isn't a picture in the corner; it's a hidden pattern woven into the recipe of the protein, detectable only by a special checker. Crucially, the marked proteins still worked normally in early lab tests.

The trick works by subtly influencing which amino acids — the protein building blocks — the AI picks. For predicted 3D shapes, DeepMind adjusted part of AlphaFold 3, its famous protein-folding model, to embed a signature directly. Researchers tested the marks on protein binders designed to latch onto specific targets, including the SARS-CoV-2 spike protein. Watermarked versions matched unmarked ones in binding strength and natural variety. They also stamped an AI-designed virus, which stayed functional.

Why does this matter? DNA-making companies currently screen customer orders against databases of known biological threats. But AI can invent proteins that look nothing like known dangers, so those lists fail — forcing slow, manual reviews. A watermark gives reviewers an extra clue: this design came from a trusted AI model. It could also keep scientific databases like UniProt and GenBank from filling up with mislabeled synthetic entries that mislead future research.

The catch: this is a proof of concept. Tests covered only a limited range of designs, and the technology only helps if companies and researchers actually adopt it. Details about the modified virus are still pending in a separate paper. As AI gets better at designing biology, tracing what it creates may become as routine as checking a receipt.