← All research storiesBiology

Peer-reviewedBiology

Scientists hid a secret “made by AI” watermark inside designed proteins, and they still worked

Google DeepMind researchers added invisible, detectable marks to AI-designed proteins and showed in lab tests that the marked proteins still worked as well as unmarked ones.

How far alongPeer-reviewedReviewed by independent experts for a journal before publication. Peer review is a quality check, not a grade of evidence strength: confidence still depends on study design, limitations, replication and the wider evidence.
How soon could it reach you?A few years outPromising, but needs more testing or engineering first. Our rough editorial guess is several years, with a lot of uncertainty.
Where it comes fromNatureGoogle DeepMind · Sep 30, 2026Read the original →

What the scientists did

Proteins are the tiny machines that do most jobs in living things. AI can now design brand-new proteins, for example ones that grab onto a virus. That’s powerful for medicine, but it also raises a question: how can anyone tell later whether a protein came from an AI tool?

The team built SynthIDBio, which slips a hidden pattern into a protein’s design. One version tweaks the choice of amino acids (the protein’s building blocks) inside a design tool called ProteinMPNN. Another adjusts AlphaFold 3 so the predicted 3D shapes carry a hidden mark. They then made hundreds of real proteins in the lab to check whether the marked ones still worked.

Words to know

Watermark

A hidden signal added to something to show where it came from, without changing how it looks or works.

  • Faint marks on paper money
  • Hidden labels some AI image tools add to their pictures

Protein binder

A protein designed to stick tightly to a specific target molecule.

  • Targets in this study: part of the COVID-19 virus’s spike protein, VEGF-A and PD-L1

False positive rate

How often a test says “yes” when the answer is really “no.”

What they found

In computer tests, the watermark on predicted protein structures was detected more than 99.8% of the time while wrongly flagging unmarked structures only 0.1% of the time. In the lab, the team made 222 unwatermarked designs and 267 watermarked designs in each of two watermarking settings, aimed at three targets. The marked binders stuck to their targets about as strongly as the unmarked ones.

>99.8%of watermarked predicted structures detected in computer tests (at 0.1% false positives)
222 + 267unwatermarked designs, plus watermarked designs in each of two settings, made in the lab
3targets the binders were tested on

Why it matters to you

You won’t see this in a store, but it’s about trust in science. As AI helps design new medicines and materials, being able to trace where a protein came from could help scientists check research claims and help keep powerful design tools from being misused.

Where this could lead

Tracing AI-designed molecules

Labs and companies could check whether a protein was made with a watermarking AI design tool.

Few years

Biosecurity checks

Watermarks could become one of several safeguards for AI tools that design biological molecules.

Few years

These are possibilities the research points toward, not promises. Most early findings take years of testing before they reach everyday life, and some never do.

Why scientists care

Watermarks are already being tried for AI-made text and images. This study shows the idea can work in biology, where even tiny changes might break how a molecule works.

Keep in mind

The mark only says “watermarked or not”; it can’t say who made the protein. Someone determined could remove it by redesigning the protein, though the authors found that tends to hurt how well it works. The structure version can be weakened by common computer simulations. It’s one safeguard, not a complete solution.

How far along is the science?

  • Peer-reviewedReviewed by independent experts for a journal before publication. Peer review is a quality check, not a grade of evidence strength: confidence still depends on study design, limitations, replication and the wider evidence.
  • PreprintA manuscript shared publicly, often before or alongside journal review. Check its separate journal and review status (each story says what we know). Results may change.
  • Conference talkPresented to other scientists at a meeting. Usually short and early; a full paper may come later.
  • Agency reportPublished by a science agency such as NASA, NOAA or NIH. Reviewed internally, but not by an outside journal.
  • Press releaseAn announcement from a university or company. Useful, but always check the research it describes.

How soon could it reach you?

  • Could reach you soonCould show up in products, services or advice soon. Our rough editorial guess is about 1 to 3 years, not a forecast from the research.
  • A few years outPromising, but needs more testing or engineering first. Our rough editorial guess is several years, with a lot of uncertainty.
  • Long-term scienceBasic research that builds knowledge. Real-world uses, if any, are far off.
  • Small stepA small update to earlier work. Useful to researchers, but not a change you would notice yet.

Every story links to its original source: the paper itself, or the conference program when no paper is out yet. Explanations are written with AI, usually from the paper’s abstract (each story says what it was written from), then checked against the source. We only use journals, recognized preprint servers, science agencies, scientific conferences and established science news outlets.