Google watermarks AI-designed proteins as Anthropic claims first AI enzyme discovery

Google's DeepMind team published a research paper on Wednesday describing a way to watermark protein sequences designed by artificial intelligence, according to Ars Technica. The system is built on Google's SynthID technology, which adds an imperceptible signal to AI-generated text and images.

It is the newest attempt to answer a question that has been open for nearly a year: how do you tell a protein that was designed by a trusted lab from one that was not, when the tools used to screen DNA sequences cannot recognise AI-generated proteins at all? The problem is one of information density. Proteins are strings of just 20 amino acids, and many of those positions are load-bearing: change the wrong residue and the protein stops working. Ars Technica reports that SynthID works by biasing the probability of certain choices the AI makes, so the bias is spread across the whole output. In an image with millions of pixels, that is easy. A 500-amino-acid protein is, by comparison, tiny.

Why proteins are hard to watermark

The risk of failure ran in both directions. The watermark might not carry enough signal to be detectable, or the model might distort the sequence so much to accommodate a watermark that the resulting protein is inactive. The DeepMind team, according to Ars Technica, tested it rather than assuming it would work. The context for that work is that the same protein design tools that have produced enzymes capable of digesting plastics or blocking venom proteins can also be pointed at toxins or viral proteins. The screening software in use does not flag AI-designed proteins, because nobody has characterised them well enough to know which ones are threats.

Anthropic's claim, and the questions around it

On 1 October, Endpoints News published a Q&A with Eric Kauderer-Abrams, Anthropic's head of life sciences, about enzyme research, the company's biology lab and a preclinical pipeline. The outlet noted that Anthropic said last week its AI system made its first scientific discovery, with Claude agents identifying a new enzyme system. That claim has not gone unchallenged. Google News listings surfaced a piece from The Scientist dated 29 September asking whether Anthropic's biology breakthrough was borrowed from another scientist's work.

The Scientist's article is not in this dossier beyond its title and URL, so the substance of that question cannot be assessed here; it is worth noting only that the priority claim is contested. The Anthropic result and the DeepMind watermarking paper are not competing solutions to the same problem, but they sit on opposite sides of a line. One is about what AI systems can produce. The other is about how anyone else can verify what they produced.

Benchmarks that measure the wrong thing

Verification is a broader problem than protein design. A paper submitted to arXiv on 29 September by Jiajun Wu, Jian Yang, Zixiang Ni, Zhenzhu Li and Bin Chong introduces CARAT, a benchmark designed to test whether materials-focused large language models reason about crystal structures or simply copy answers printed in their inputs. The authors' point is that accuracy cannot distinguish the two. A structural description often prints the very field it is being scored against, so a model can score well by repeating what it was shown. CARAT holds question and gold answer fixed across eight matched views and names each structural relation separately in what the authors call GraphSpace.

The results include a self-critique. On the hardest families, the grounded view is worth 17.3 points over formula inputs, the paper says. GraphSpace beats a plain periodic graph by 19.3 points, but the authors break that margin down: where the plain rendering already contains everything the question needs, the gain is 1.96 points, and where it omits those fields entirely, the gain is 46.7 points. Their conclusion is that the headline number mostly measures what the baseline lacked.

The paper also attacks its own benchmark. A rule that skips a link and reads the list directly answered four of seven hardened families, so the authors rebuilt the benchmark until eleven such shortcuts sat near chance. On a frozen model, the system quoted a link but returned the same answer when the link was redirected in 95.6% of paired cases: it repeated the relation without using it. After matched supervision the figure reached 99.8%, and deleting the link dropped performance to 23.4%, below the 27.0% achieved by the best shortcut.

The paper's own summary is blunt: the frozen model repeats the relation without using it, on 95.6% of paired cases.

That is a useful corrective at a moment when AI-for-science results are arriving faster than they can be checked. A model that recites a crystal structure and a model that reasons from one look identical on a leaderboard, in the same way that an AI-designed protein and a natural one look identical to a DNA screener.

Background: what is moving in the field

Commercial activity around AI-driven protein and materials design has continued through the same period. On 1 October, PA Media reported that Hansa Biopharma is partnering with Cradle to advance AI-driven protein design and engineering. Bioengineer.org reported on 30 September that an AI-guided recipe had tamed the chaos of high-entropy perovskite solar materials. R&D World reported the same day that its parent company Arrowfly launched AI for Engineers, a platform aimed at engineers navigating AI.

On the research side, Nature published work on 28 September describing accelerated discovery of thermostable mRNA-lipid nanoparticle vaccines using data-efficient AI, according to a Google News listing. And on 1 October, Stanford HAI announced 15 new data science scholars. None of these are the same kind of claim as a watermarking scheme or a reasoning benchmark, but they explain why the verification problem matters. Protein design and materials discovery are being pushed forward by a widening set of labs, startups and pharmaceutical companies. The tools for checking their output are not moving at the same speed.

For its part, the DeepMind watermarking work, as Ars Technica describes it, is not a screening tool. It does not identify dangerous proteins. It identifies which proteins came from a trusted designer, so that everything else can be treated with more suspicion. That is a narrower claim than it may first appear, and it depends on designers choosing to use the watermark at all.

Older context

The dossier also includes a ScienceDaily report from 25 September, based on a University of British Columbia study published in Organisms Diversity & Evolution, describing two new sea spider species from the Salish Sea. Lead author Cormac Toler-Scott said he was interested in how sea spiders are affected by climate change and human disturbance, and could not find existing work. The study, which collected specimens between September 2023 and August 2024 at depths up to 18 metres, marks the first new sea spider species formally described from the Salish Sea in nearly 100 years.

It is not an AI story. It is a reminder of the baseline problem that runs through all of the above: the species had been there, and nobody had described them. The same is true of AI-designed proteins. They exist, they are being made, and until recently there was no agreed way to mark them.