Tag

#SynthID

SynthID is Google DeepMind's attempt to watermark AI-generated content in a way that survives compression, cropping, and the slow degradation of the internet, and these posts follow it with close attention. I write about the technical approach, embedding statistical signals in generated text, audio, and images that are imperceptible to humans but detectable by a trained classifier, and about the open questions it raises. Can a watermark survive an adversary who wants it gone? Should provenance be invisible or declared? Who holds the detector and who trusts them? The pieces here treat SynthID as a serious technical contribution and as a political artifact, because any watermarking standard that becomes default is also a standard of evidence.