Google’s deepfake detection system passed a real-world test this week. A fabricated image of Kentucky Senator Mitch McConnell, shown covered in tubes and in visible distress in a hospital bed, spread fast across Reddit and X before fact-checkers caught it. According to TechCrunch AI, the fake was exposed because it carried SynthID, Google’s invisible watermark for AI-generated images. By Wednesday, Snopes had debunked the picture, noting that the image registered as containing the SynthID signature. The watermark did exactly what it was built to do.
This matters because it’s one of the first high-profile cases where anti-deepfake tech actually worked in the wild, not just in a lab demo.
What SynthID does
Google launched SynthID at its I/O developer conference in 2025. Think of it as a signature baked into the pixels of an AI-generated image. It’s invisible to you and me, but SynthID’s algorithms can read it. The clever part: because the mark lives inside the image itself, it survives screenshots and reposts across platforms. The McConnell fake got screencaptured and passed around repeatedly, and the watermark still held up. That durability is the whole point, and it’s where most detection methods fall apart.
Why the timing made it dangerous
McConnell’s health has been under a microscope. He checked into the hospital after an emergency call on June 14 and has mostly stayed out of public view since, according to TechCrunch AI’s reporting. That silence created a vacuum, and a convincing fake image dropped right into it. When speculation is already running hot, a realistic photo doesn’t need much help to go viral. People want confirmation of what they already suspect. This is precisely the scenario where deepfakes do the most damage, and precisely why a working detection layer matters.
The catch: it only works if the tool opts in
Here’s the real limitation. SynthID can only flag an image if the tool that generated it participates in the program. No participation, no watermark, no detection. The current lineup:
- Google Gemini models have carried the watermark since launch in 2025.
- OpenAI joined in May 2026, part of a wider push against malicious image generation.
- Anthropic does not participate.
So the coverage is real but partial. A fake made with a non-participating tool, or an open-source model, sails through clean. What stands out here is that the McConnell image happened to be made with a participating tool. That’s a win, but it’s also a reminder of how much slips outside the net.
How you can check an image yourself
You don’t need special access to test this. Two options right now:
- Ask a Gemini model whether an image contains the SynthID watermark.
- Upload the image to OpenAI’s public image verification tool.
For journalists, moderators, and anyone doing quick verification, that’s a genuinely useful pair of checks. Free, fast, and no technical setup.
Why this is a bigger deal than one photo
The watermark-and-detect approach has faced plenty of doubt. Critics argued it wouldn’t survive real-world sharing or that adoption would stay too thin to matter. This case answers the first worry directly: the mark survived multiple reposts and a screencapture, and a major fact-checking site used it to kill a viral hoax. That’s the proof-of-concept the technology needed.
The second worry still stands. As long as participation is voluntary and major players sit out, bad actors have easy workarounds. The system catches the honest tools, not the ones trying to deceive you. Watermarking is a floor, not a ceiling.
Still, this is progress you can point to. A year ago, debunking a fake like this leaned on visual analysis, source-tracing, and gut instinct. Now there’s a machine-readable signal that fact-checkers can query in seconds. Expect more platforms to feel pressure to join, and expect verification tools to get baked into the places where images spread. The gaps are real, but the direction is clear. For the full breakdown, check the original report at TechCrunch AI.