Google's DeepMind team released protein watermarking technology on Wednesday, saying it allows for the identification of proteins designed by trusted researchers without compromising their function. It is the company's first biosecurity update since the risk of AI-designed proteins was flagged nearly a year ago.
Google reported the system works by embedding a watermark in protein sequences, measured on the effectiveness of its detection process. That compares with the prior lack of a reliable method for identifying AI-designed proteins.
The technology is built on Google's SynthID platform and targets biosecurity applications. Availability begins with research publications, initially for academic and biotech communities.
"The watermark influences the probability of certain choices the AI makes, and that bias ends up systematically distributed throughout the product," said David Baker, a researcher involved in the project. The system allows for the detection of watermarked proteins through a key-based scanning process.
The announcement follows a growing concern over AI's role in biological design. Google's approach is seen as a step toward managing potential threats while enabling innovation.
Google did not say how the system will be distributed or maintained, and it raised questions about the effectiveness of the watermarking in very short proteins. The team highlighted the need for further testing and refinement before widespread use.
Source: arstechnica
