Google's SynthID watermark is difficult to remove, but it doesn't solve the problem of AI-driven disinformation.
Google/DeepMind
OpenAI
Runway
NVIDIA
DeepMind
A research collaboration between the Starling Lab at Stanford and the University of Southern California has calculated that up to 1975, humanity created 1.5 billion images over 149 years, while AI generated the same number in just 18 months. Google is rolling out SynthID, a watermark to label AI-generated content, which the company claims is resistant to editing, but experts note that it does not fully address the problem of disinformation.
The scale of AI-generated content is enormous: Starling Lab estimates that humanity created 1.5 billion images from 1826 to 1975, while generative AI did it in 18 months. At the I/O conference, Google announced that its tools have generated over 100 billion images and videos over several years. To label AI content, Google uses the SynthID watermark, which the company says remains visible even after compression, editing, and other modifications. The watermark is embedded into pixels or audio waves, unlike C2PA metadata, which is easy to remove. Google DeepMind scientist Pushmeet Kohli stated that SynthID is designed as a robust way to label content. However, despite this, the system does not solve the problem of AI disinformation, as the watermark can be circumvented.
- Сокращения
- C2PA = Coalition for Content Provenance and Authenticity — Коалиция за происхождение и подлинность контента
Source: Ars Technica —
original