SynthID: How Google Is Watermarking AI-Generated Content
SynthID, covered here https://deepmind.google/discover/blog/identifying-ai-generated-images-with-synthid/, addresses a problem that only gets harder as AI generation quality improves: how do you know if an image was generated by AI rather than captured by a camera?
The watermarking approach being imperceptible to human viewers but detectable algorithmically is the technical design that makes SynthID practically useful rather than just theoretically sound. A visible watermark would be trivially removed. An invisible signal embedded in the image at the pixel level during generation survives most editing and compression operations without affecting visual quality.
The honest limitations worth acknowledging: SynthID only identifies content generated by Google's own tools. Content generated by other systems, AI output that passes through multiple editing steps, and screenshots of AI content all reduce the detection reliability. A watermarking system that only covers one provider's output is useful but not a comprehensive solution to AI media provenance.
Whether watermarking or media literacy is the more tractable approach to AI misinformation is the genuine policy debate. Technical solutions have the advantage of working without requiring behavioural change. Media literacy has the advantage of being generalisable rather than specific to particular generation systems.
Can watermarking realistically solve the AI misinformation problem, or does media literacy matter more in the long run?