Earlier this week, a photo started making the rounds on Reddit and X. It showed Kentucky Senator Mitch McConnell in a hospital bed, covered in tubes, looking like he was in pretty bad shape. Given that McConnell has been out of the public eye since a hospital visit back in June, the image felt plausible to a lot of people. The kind of thing that makes you stop scrolling and go “wait, is this real?”
Turns out, it wasn’t. And the technology that busted it is worth talking about — not just for this one hoax, but for what it says about where we’re headed.
The Watermark That Worked
Fact-checking site Snopes picked up the image and ran it through Google’s SynthID detection system. The result came back clear: the image carried a SynthID watermark, meaning it was generated by an AI tool that participates in Google’s watermarking program.
And here’s the thing — that watermark isn’t something you can see with your naked eye. It’s not a little logo in the corner or a subtle pattern you can spot if you squint. It’s embedded into the image data itself at the pixel level, invisible to humans but easily spotted by the right detection tools. And because it’s baked into the actual image content rather than just attached as file metadata, it survives screenshots, re-uploads, and compression.
The image was shared widely across Reddit and reshared on X, but the watermark traveled with it each time.
How SynthID Actually Works
Let me break this down because it’s genuinely clever engineering.
SynthID works by embedding an imperceptible digital signature into AI-generated content at the moment of creation. It’s not an afterthought or a post-processing step — the watermark is part of the generation process itself. For images, this means modifying pixel values in a way that’s invisible to humans but creates a detectable statistical pattern.
Google DeepMind launched the system at I/O 2025, and it’s been evolving since. The watermarking covers images, audio, video, and even text. For images specifically, the watermark is designed to survive common transformations — cropping, adding filters, changing frame rates, lossy compression. The McConnell image went through multiple sharing cycles on different platforms and the watermark still held up.
This isn’t the only system in town. OpenAI joined the SynthID program in May 2026, adding both SynthID watermarks and C2PA Content Credentials to images generated through ChatGPT, Codex, and their API. C2PA is a separate — but complementary — approach that stores cryptographic provenance metadata alongside the image. Think of SynthID as the tattoo that stays with the content and C2PA as the passport that tells you where it came from. Together they’re more powerful than either alone.
Anthropic doesn’t participate in SynthID yet, which means if someone generates a deepfake with Claude, there’s no SynthID watermark to catch it.
Why This Specific Case Matters
I’ve written before about how Reddit is fighting AI spam with AI — the arms race between AI-generated content and the tools designed to detect it. But the McConnell hoax is a different category entirely. It’s not spam, it’s not a marketing gimmick. It’s political disinformation, targeted at an active U.S. Senator, designed to spread during a genuine health crisis.
What makes this case notable is that the detection system worked exactly as designed. That’s still rare enough to be newsworthy. Most of the time, deepfakes fly under the radar. They hit WhatsApp groups and Facebook pages, get shared by well-meaning relatives, and by the time anyone debunks them, the damage is done.
This time, Snopes caught it fast, SynthID confirmed it, and the story moved from “is McConnell dying?” to “someone generated a fake image of McConnell.” That’s the ideal flow — detect, verify, contextualize — happening in public.
The Gap in the Safety Net
But let’s be honest about the limitations.
SynthID only works when the AI tool that generated the image chooses to participate in the watermarking program. If someone generates a fake image using a non-participating model — say, a locally-run Stable Diffusion instance or a tool from a developer who hasn’t joined the program — there’s no SynthID watermark to detect. The image could be just as convincing, and no amount of Google technology would flag it.
This is the same challenge I covered in my piece about Midjourney wanting Hollywood to come clean on AI — the transparency problem only works when everyone plays by the same rules. SynthID is a voluntary program. That’s not a knock on Google — building an opt-in system at this scale is genuinely impressive. But voluntary compliance leaves gaps, and bad actors will exploit those gaps.
There’s also the question of access. The SynthID Detector portal is currently invite-only for journalists and media professionals. That means the average person scrolling through social media can’t easily check if an image is watermarked. You can upload suspicious images to Gemini and ask it to check, which is more accessible, but it requires knowing that the option exists in the first place.
What This Means Going Forward
The McConnell hoax is a perfect stress test for where we are with AI content authentication in 2026. It tested the system against a real-world scenario: generated image, multi-platform sharing, high-stakes subject matter, time-sensitive debunking needed. The system passed, but barely — and only because Snopes was doing the detective work.
I’m reminded of what happened with Discord’s AI moderation failures where the algorithms just weren’t ready for the nuance of human behavior. Detection technology is in a similar place — promising, but not yet trustworthy enough to be the sole line of defense.
What would make this stronger? Three things:
- Wider adoption. If every major AI image generator participated in SynthID or a compatible standard, the baseline for detection would be much higher. OpenAI joining was a big step, but we need Anthropic, Midjourney, Stability AI, and others on board too.
- Public-facing detection tools. The SynthID Detector portal needs to be open to everyone, not just journalists. Right now, the people most likely to encounter deepfakes — regular social media users — have the least access to verification tools.
- Combined approaches. C2PA and SynthID are complementary, not competing. Metadata tells you origin, watermarks tell you authenticity. Using both gives you a much fuller picture than either alone.
The Bigger Picture
I’ve been thinking a lot about trust in the AI era — how we decide what to believe when the tools for creating convincing fakes are available to anyone with a decent GPU. It’s not a theoretical question anymore. The McConnell image wasn’t generated by a state-sponsored disinformation lab. It was shared by regular accounts on Reddit and reshared by regular people on X.
And that’s the part that keeps me up at night. The barrier to entry for creating political deepfakes just keeps dropping. A year ago, generating a convincing fake of a public figure required technical skill and expensive hardware. Now? Someone with a Gemini subscription or a ChatGPT Plus account can do it in seconds. The technology for detection has to keep pace with the technology for generation, and right now, it feels like detection is running a step behind.
But moments like this — where the watermark works, the facts get checked, and the hoax gets exposed before it causes real damage — those are worth celebrating. They prove the model works. They just prove we need more of it.