Google's SynthID Watermark Helps Debunk a Viral McConnell Deepfake

A fake image of the ailing senator on life support spread across X and Reddit — until fact-checkers found Google's invisible AI watermark baked into it, in a rare public win for deepfake detection.

Jul 9, 2026Source: TechCrunch

Google's SynthID Watermark Helps Debunk a Viral McConnell Deepfake

For once, the watermark did its job. A fabricated image of Senator Mitch McConnell lying in a hospital bed "covered in tubes… in a state of extreme distress" spread across X and Reddit after a July 7 post, feeding speculation about the senator's health. By the next day, fact-checkers had shot it down — using Google's SynthID watermark hidden inside the picture itself.

How it was caught

Snopes found the image "registers as containing the SynthID watermark designed by Google to identify AI-generated pictures." Lead Stories independently confirmed it, running the picture through a public verification tool that flagged the watermark and indicated the image was generated with AI — backed up by tell-tale artifacts like tangled tubing connected to nothing, waxy skin and warped background equipment. The detection was done by fact-checkers using public tools, not by Google itself.

The twist

The revealing detail: the image appears to have been made with OpenAI's tools, yet it still carried Google's watermark. That's because OpenAI joined the SynthID program in May 2026, so images from its generators now embed the signal — a rare instance of rival labs' safety infrastructure interlocking to catch a fake neither would want spread.

Why the watermark survived

SynthID embeds an invisible, pixel-level signature readable only by its detection algorithms and imperceptible to viewers. Crucially, because it's baked into the image rather than attached as metadata, it survives screenshots and re-sharing — exactly the conditions under which the McConnell picture bounced across platforms. Metadata-based provenance systems like C2PA, by contrast, are trivially stripped by a screen capture.

The catch

The limitation is opt-in coverage. SynthID only marks images from tools that participate — Google's Gemini has watermarked since 2025, and OpenAI, Nvidia and ElevenLabs have since joined — but Anthropic does not participate, and open-source models produce unmarked images. The Next Web called this McConnell debunk SynthID's first high-profile catch, "a rare but significant win." The context that made it matter: McConnell was hospitalized after an emergency on June 14 and has been largely out of public view, leaving an information vacuum that a convincing fake rushed to fill.

Why It Matters

This is the first time AI watermarking visibly worked at the speed of a real viral news cycle — evidence that invisible provenance signals, not after-the-fact forensics, may be the practical line of defense against deepfakes. But the opt-in gap is the whole story: as long as Anthropic and open-source generators stay outside the system, watermarking catches the careless, not the determined, and the next fake can simply be made with an untagged tool.

Filed underGoogle AI AI Safety #Google #SynthID #deepfakes #misinformation