How Google's Watermark System Exposed a Fake Hospital Photo of a Senator

Snopes debunked a viral image purporting to show Senator Mitch McConnell in a hospital bed by detecting an invisible marker that Google embeds into AI-generated pictures TechCrunch. The image had spread across Reddit and X before fact-checkers flagged it as synthetic Snopes, and TechCrunch reported on the debunking July 8.

McConnell entered a hospital on June 14 after an emergency call and had stayed largely out of the public eye since. That absence created an opening for the fabricated image: a senator already known to have health challenges, no recent sightings, and a dramatic photo ready to be believed at first glance. Snopes did not identify who created the image or which AI tool they used; the fact-check hinged on the technical detection method itself.

What SynthID Does

SynthID is Google's watermarking system, launched at the company's developer conference in May 2025. It plants an invisible signal into AI-generated images at the pixel level—think of it as a fingerprint woven into the image data itself. The watermark is designed to survive the wear and tear of being copied, compressed, and reshared across the internet, which proved useful in this case: the McConnell image had already been reposted and recompressed many times before fact-checkers examined it. Since Gemini (Google's AI assistant) started embedding the watermark, anyone can upload an image to Gemini or OpenAI's verification tool and check whether it carries the signature.

Google framed SynthID not as a consumer product but as a foundation layer for verifying where AI images come from—a standard piece of infrastructure for a world where AI-generated pictures will be common.

Growing Adoption, But Gaps Remain

OpenAI joined the SynthID program in May 2026 and now runs its own public verification tool, extending the watermark check beyond Google's ecosystem TechCrunch. Anthropic has not yet joined.

That absence matters. Watermark-based detection only works if the AI system that generated an image actually embedded the watermark in the first place. If someone uses a tool from a lab that doesn't participate in SynthID, or runs the image through an open-weights model (freely available software) that strips or never applies the watermark, the verification system has no signal to detect. The network effect cuts both ways: the more AI labs that sign up, the more reliable the system becomes; holdouts create blind spots.

A Shift in How Fact-Checkers Work

The broader significance of this case hinges less on the McConnell image itself—it was a fairly standard political hoax—and more on the fact that a major fact-checking outlet now uses a cryptographic watermark check to debunk a synthetic image rather than relying solely on visual forensics or source-tracing. That represents a methodological shift in the fact-checking world, even if this particular incident was modest in scope.

The real test for SynthID and similar systems will arrive when someone manufactures a hoax specifically designed to evade detection—either by using a non-participating AI model or by running the image through a pipeline engineered to strip metadata and watermarks. That test has not yet reached public notice, at least not in a case that made headlines.

Why This Matters for Politics

Political deepfakes involving health status carry weight because they exploit the exact kind of information gap that McConnell's hospitalization created. A senator absent from public view for weeks becomes a blank canvas; an AI-generated hospital image can fill it with plausible-seeming claims, sympathetic or damaging. SynthID does not prevent someone from creating such an image. What it does is give fact-checkers—and increasingly, ordinary users with access to Gemini or OpenAI's tool—a way to verify one before sharing it further. Whether that check happens quickly enough, before the image has already spread too far to correct, remains an open question, and one that will likely repeat itself in future cases.

Worth flagging: A watermark detection tells you an image came from a participating AI model. By itself, the absence of a watermark does not prove an image is authentic. It could have come from an older model, a non-participating lab, or a source that has nothing to do with AI at all—a straightforward photograph, for instance. The reliability of watermark-based detection depends on how widely these systems are adopted across the AI industry.