The watermark worked. That's the headline.

A fabricated photograph of Senator Mitch McConnell, tubes snaking across his face, distress etched into every pixel, raced across Reddit and X earlier this week. By Wednesday, Snopes had killed it. The image carried Google's SynthID signature — an invisible stamp the company's own algorithms can read but human eyes cannot see. The system did exactly what it was designed to do: catch a lie before it hardened into consensus.

This matters because deepfake detection usually fails in public. Researchers publish benchmarks. Academics cite precision-recall curves. But a live, high-stakes debunking on a sitting senator's health? That's the first time the plumbing has held under pressure.

McConnell has been absent since an emergency hospitalization on June 14. The vacuum invited invention. Someone — unknown, unaccountable — fed a prompt into a model and released a picture that looked like evidence. The internet did what it does: shared first, asked questions never. Without SynthID, the image might still be circulating as "proof" of a cover-up.

Google launched SynthID at I/O 2025. The architecture is simple in theory, stubborn in practice. The watermark lives inside the pixel data. Screenshot it, re-compress it, upload it through three different platforms — the signature survives. That resilience matters. Most provenance schemes break the moment an image leaves its native app.

But the system has a structural flaw it cannot code around: participation is voluntary.

Gemini models have carried the watermark since launch. OpenAI joined in May 2026, part of a broader safety push the company now treats as brand infrastructure. Anthropic has not. That means every image Claude generates — and every image generated by any model Anthropic licenses or influences — moves through the world unmarked. Invisible. Uncheckable.

Google knows this. So does OpenAI. So does every policymaker who has sat through a briefing on synthetic media. The detection layer only covers the generators willing to be detected. The rest operate in shadow, and the shadow is growing.

Users can verify images two ways now: ask a Gemini model, or upload to OpenAI's public verification tool. Both work. Both are free. Both require the user to suspect forgery in the first place. That's a behavioral dependency no protocol can eliminate.

The McConnell hoax didn't fail because the watermark was clever. It failed because the generator that made it participated in the program. If the same prompt had gone to a non-participating model, the image would have no signature. Snopes would have had nothing but visual analysis — eyes, context, guesswork. The debunking would have been slower, softer, contestable.

That's the real story. Not that SynthID worked this time. That it only works when the maker agrees to be caught.

Anthropic's refusal isn't technical. It's strategic. The company has positioned itself as the safety-first alternative, yet it declines the one interoperable standard that lets the public verify its outputs. That contradiction deserves scrutiny every time Anthropic publishes a safety paper or testifies about responsible deployment.

Legislators have noticed. The Senate Commerce Committee's draft AI transparency bill includes a provision that would mandate watermark participation for any model deployed commercially in the U.S. Industry lobbyists are already calling it "premature standardization." They're right that the tech is young. They're wrong that the alternative — voluntary adoption — is acceptable.

Voluntary adoption gave us this week's result: one caught hoax, an unknown number of uncaught ones. The asymmetric risk is obvious. A single undetected deepfake of a presidential candidate in October 2028 would cost more than the entire compliance burden of mandatory watermarking.

Google deserves credit for building a system that survived its first real test. OpenAI deserves credit for joining it. But the ecosystem cannot run on credit. It runs on coverage.

Until every major generator participates, SynthID is a net with holes the size of Anthropic's market share. The McConnell image proved the net holds. It didn't prove the net covers the water.

Next time the senator's health becomes a vacuum, the forger will choose a different model. Count on it.