tech
Google's SynthID watermark is hard to break, but it doesn't solve AI misinformation
Deciding what's real on the Internet won't be easy in the future.
TL;DR
- Generative AI has produced an immense quantity of media, with Google alone creating over 100 billion AI images and videos in a couple of years.
- Google's SynthID is an invisible watermarking technology designed to label AI-generated content, intended to be robust against edits and data loss.
- Testing showed SynthID surviving hundreds of simulated degradation cycles and even screenshots, though it could be defeated by significant cropping.
- Challenges remain with SynthID, including limited access, fragmentation across different AI platforms, and the potential for adversarial attacks.
- The proliferation of open-source AI models means that much AI content will likely remain unlabeled, rendering watermarks insufficient as a sole solution.
- The article suggests that focusing on verifying authentic content through technologies like C2PA might be more effective than solely attempting to label AI-generated material.