NemoClaw Knowledge Wiki

AI Watermarking: Distinguishing AI-Generated Content and Biological Sequences

4 min read

🗂️ Creative Pursuits · View mindmap

AI Watermarking: Distinguishing AI-Generated Content and Biological Sequences

Clip title: From deepfakes to DNA: the science of watermarking AI Author / channel: Google DeepMind URL: https://www.youtube.com/watch?v=HIUzrxQxTtw

Summary

This video from Google DeepMind addresses the pressing challenge of distinguishing AI-generated content from human-made or natural content, a concern growing rapidly with advancements in generative AI across various modalities. Professor Hannah Fry introduces the core problem: as AI becomes increasingly adept at creating realistic text, images, audio, video, and even biological molecules, the fundamental question “Is that real?” becomes profoundly difficult to answer. This raises significant risks, from the spread of misinformation to the potential design of harmful biological structures. The proposed solution is a sophisticated form of watermarking, not a visible logo, but a subtle, mathematical, and invisibly embedded signal.

The discussion highlights Google DeepMind’s pioneering work with SynthID, a watermarking system. Pushmeet Kohli, an architect of SynthID, explains that the primary motivation is to provide a clear sense of provenance, allowing users to verify the origin of digital content. Jeremy Ratcliff details the application of this technology to biology with “SynthID Bio,” aiming to differentiate natural biological sequences from those created by AI, especially given the potential for AI to design molecules with problematic properties that are not immediately apparent from their sequence alone. A good watermark, they explain, must be imperceptible to humans, highly detectable by systems, robust against attempts to remove or alter it, scalable across different data types, and must not degrade the quality or function of the content.

The guests elaborate on how watermarking works in practice. For text, large language models generate words based on probability distributions, and SynthID subtly biases these choices using a secret key, embedding patterns that a detector can later recognize. For images and video, a neural network subtly modifies the content in an imperceptible way, co-trained with a detector. An adversarial agent constantly tries to erase the watermark by introducing changes like scaling or noise, ensuring the system’s robustness. This technology is already being used; Google’s Gemini app can detect AI-generated content from partnered sources, playing a crucial role in combating misinformation. In biology, SynthID Bio applies similar principles to protein sequences and 3D structures, ensuring that AI-designed proteins maintain their intended function while carrying an invisible tag of their artificial origin.

The implications of this technology are far-reaching. SynthID Bio directly addresses the biosecurity concern that AI could generate novel, harmful biological molecules that current DNA synthesis screening systems might miss due to their unconventional sequences. By watermarking AI-generated proteins, an additional safeguard is put in place, allowing synthesis companies to identify the artificial origin and exercise increased caution. Experimental results have shown that watermarked proteins, even when physically synthesized in a lab, retain their intended function without degradation. While challenges remain, including the need for industry-wide standards and continuous adaptation in a “cat-and-mouse” game with malicious actors, Google DeepMind’s efforts to open-source this technology and foster collaboration aim to establish a universal “certificate of authenticity” for AI-generated content. The ultimate goal is to proactively harness the immense power of AI for societal good, such as drug discovery and human health, while preventing its misuse before threats fully materialize.

Video Description & Links

Description

How do you know if what you’re seeing is real? Professor Hannah Fry sits down with Pushmeet Kohli (VP of Science) and Jeremy Ratcliffe (Research Scientist) to break down SynthID, the watermarking technology expanding from text, images, and video straight into biology. Learn how invisible, mathematical watermarks help trace AI provenance, protect against misuse, and safeguard protein sequence design without impacting biological function.

Timecodes: 00:00 Introduction 00:34 What is a watermark? 04:35 SynthID 15:16 Images and video 19:28 SynthID Bio 28:00 Future of resilience

Learn more about SynthID Bio: https://deepmind.google/blog/introducing-synthid-bio/

URLs

Graph View