The watermark holds. The trust layer doesn't.

Detection has outrun verification. Watermarking technology has arrived that genuinely holds up: Google's SynthID, an invisible watermark baked into the pixels of AI-generated images and video, survives the compression, resizing, re-edits, and well-worn-meme recirculation that define how images actually travel online. That is real engineering progress, and it solves only the first half of the problem.

A durable mark on a file is not a provenance system. Provenance is the chain that says where the file came from, who vouched for it, and whether the label still travels with it when somebody takes a screenshot or re-saves the image. The Coalition for Content Provenance and Authenticity (C2PA), a cross-industry metadata standard, is cryptographically secure and trivially strippable; pixel-embedded watermarks survive that and tell you nothing about who vouched. Two labels, one unbuilt trust layer.

The 100 billion AI images and videos Ars Technica reports Google has produced sit exactly in that gap. Detection capability is racing ahead of the institutions and platforms that would have to read the label, believe it, and act on it. Until that catches up, the answer to "is this real" still belongs to newsrooms, platforms, and the reader holding the screen.