GoKawiil
Tech News
Google's DeepMind published research describing a method to embed watermarks directly into AI-designed protein sequences without disrupting their biological function. The system, adapted from Google's SynthID technology used for digital media, aims to let trusted researchers' AI-generated proteins be verified while flagging unidentified sequences for further scrutiny.
Google DeepMind researchers in London have built a watermarking system, described in Nature, that embeds a hidden statistical signal into both the amino-acid sequence and 3D structure of AI-designed proteins. The tag, called SynthIDBio, is meant to mark proteins as machine-generated without impairing their biological function, allowing them to be identified later in databases.
Researchers have extended Google's SynthID-text watermarking system to ProteinMPNN, a widely used protein sequence design model, creating a method called SynthIDBio-sequence. The approach embeds a detectable signal during the autoregressive sampling of amino acids, using a random seed, scoring function and tournament sampling algorithm similar to the text version. Because protein design typically uses low-temperature sampling that limits randomness, the team also developed a distortionary variant with an added filter to preserve detectability.
A new report highlights that Google's SynthID and comparable systems from OpenAI and other tech firms embed invisible signals into images, audio, text and video that go beyond simple authenticity marks. These 'spymarks' can carry encoded database identifiers—potentially linkable to a user's name, IP address, birthdate or political affiliation—rather than just flagging content as AI-generated. Research on SynthID-Image shows a 512x512 image can carry a 136-bit payload, enough for a 64-bit unique identifier plus error correction.
New research from Lasso Security found that SynthID-Text, the watermarking scheme Google open-sourced and Anthropic plans to adopt for future Claude models, does more than mark AI output as machine-generated. It also changes which tools a model calls and how likely it is to follow or break its own safety rules, especially when facing adversarial prompts designed to extract sensitive data.
Anthropic has released a web page at claude.com/check-content that lets users upload files to see if Claude was involved in creating them, using embedded content credentials under the C2PA standard. The checker currently supports only multimedia formats like JPEGs, GIFs, MP3s and MOVs, not text documents. A separate text-detection API exists but remains in private preview for select organizations.