On July 7, 2026, an image started circulating on Reddit and X that showed Kentucky Senator Mitch McConnell, 84, lying in a hospital bed surrounded by a tangle of tubes and wires, looking gravely ill. One Reddit user summed up the reaction: "Oof madone! He looks terrible." The image spread fast, arriving three weeks into a real hospitalization the senator's office had explained almost nothing about.It was fake. And the way it got caught is the more interesting story.
Fact-checkers at Snopes ran the image through OpenAI's verification tool and found something the people who made the image probably didn't know was there: a hidden SynthID watermark, the same invisible signature Google embeds in content generated by its own AI models. The watermark had survived being posted, screenshotted, and reposted across two platforms, and it was enough to settle the question. The image wasn't a leaked hospital photo. It was generated.
Three days later, a second fake made the rounds: a fabricated screenshot claiming McConnell had posted a "proof of life" photo on X, holding up a copy of The Washington Post dated July 7. Then, on July 12, McConnell's office released what appears to be a genuine photo of the senator and his wife, Elaine Chao, in his hospital room, holding that Sunday's Washington Post sports section. Social media users immediately began picking apart the real photo for AI artifacts too, scrutinizing his wrist, the newspaper text, and the lighting, insisting it was fake as well. Late night host Jimmy Kimmel added to the confusion on purpose, posting his own AI-generated parody of the same photo to Instagram with himself swapped in for the senator.Three images, three days, one very confused internet. It's a tidy case study in what happens when a real news event collides with generative AI, and it's worth walking through what actually distinguishes a real photo from a fake one right now, because the tools and the confusion are both here to stay.
What made the first image so convincing
The McConnell hospital hoax worked because it had a foundation of truth to stand on. McConnell really was hospitalized starting June 14, 2026, following what his office later described as a fall. Details were scarce for weeks, which is exactly the kind of information vacuum that makes a fabricated image plausible. Nobody had a real photo to compare it against, so a convincing fake filled the gap.This is a pattern worth remembering: deepfakes rarely invent a story from nothing. They attach themselves to a true event and fill in the parts nobody has confirmed yet. The technique works as well for corporate crises and market-moving rumors as it does for a senator's hospital stay.
How SynthID actually catches these images
SynthID, developed by Google DeepMind, doesn't work like a visible watermark or a logo stamped in the corner. It's embedded during the generation process itself. When a diffusion model like Imagen creates an image, it works in a compressed "latent space" before rendering final pixels, and SynthID modifies that latent representation with noise patterns that are statistically distinguishable from random noise but invisible to the human eye.A few things make this different from older watermarking approaches:
- It's baked into the pixels, not the metadata. Traditional watermarks live in EXIF data or file headers, which get stripped the moment someone takes a screenshot. SynthID's signal survives screenshots because it's part of the image content itself.
- It's spread across the whole image. The watermark pattern repeats across multiple regions and frequency bands, so cropping or partially editing an image doesn't necessarily kill it. Detectors can often reconstruct enough of the pattern from what's left, similar to how error correction works in a QR code.
- Detection gives you a confidence score, not a yes/no. Rather than a binary verdict, tools report something closer to "80% confident this contains a SynthID watermark," which matters because degraded or heavily compressed images produce weaker signals.
OpenAI adopted the same standard in May 2026, meaning images generated through ChatGPT, DALL-E, and the OpenAI API now carry SynthID watermarks too. That's why an OpenAI verification tool was able to flag an image that, based on visual style, looked like it may have originated from a different pipeline. The two biggest image generation companies now share a common detection layer, which is a bigger deal than it sounds. It means a fact-checker doesn't need six different tools for six different AI models, at least for now.
Where the detection story gets messy
The McConnell case is a genuine win for watermarking, but it's not proof the problem is solved. A few limitations are worth being honest about.First, SynthID only detects content made by models that participate in the watermarking program. Stable Diffusion, Midjourney, and plenty of open-weight models aren't part of it, so an image generated through those tools won't carry a SynthID signature at all, watermarked or not. Absence of a watermark doesn't mean an image is real. It might just mean it came from a tool that doesn't watermark.Second, cross-tool detection is inconsistent. One tech reporter who tested this directly found that Google's own detector correctly identified an image made with Google's model, but failed to catch one generated by ChatGPT. OpenAI's Verify tool caught its own original output but missed a screenshot of that same image, and also failed to recognize an image generated by Gemini. Small sample size, but it tracks with what you'd expect: cross-platform, cross-model detection isn't a solved problem, it's a patchwork of overlapping systems that don't fully talk to each other yet.Third, and this is the part that made the McConnell story genuinely funny in a dark way, watermark detection can't rescue you from human confirmation bias. When McConnell's office released what appears to be an authentic photo on July 12, people who had just spent five days learning to distrust images immediately turned that suspicion on the real one. Newsweek asked two different AI chatbots, Grok and Copilot, to assess the same real photo. Grok said it showed signs of AI generation. Copilot said it saw no strong evidence of that. Same image, two AI assistants, two different answers. If you're using a general-purpose chatbot as your deepfake detector, you should expect exactly this kind of inconsistency, because those tools weren't built for forensic image analysis in the first place.
How to actually check if an image is AI-generated
If you find yourself needing to verify an image today, here's a more reliable order of operations than asking a chatbot for a vibe check:
- Check for a known watermark first. Google's SynthID Detector portal and OpenAI's image verification tool are purpose-built for this and will tell you whether a specific, known watermark is present. They're free and don't require an account.
- Don't trust a "no" from a watermark detector as proof of authenticity. It only tells you the image wasn't made by a model in that specific watermarking program. A clean scan from SynthID's detector doesn't rule out Stable Diffusion, Midjourney, or a dozen other tools.
- Look for corroboration from a source with something to lose. In the McConnell case, the strongest evidence wasn't pixel analysis at all, it was the fact that no credible news outlet using Google News or Yahoo News search could find any report matching the "proof of life" screenshot claim. A story that only exists on one anonymous account and nowhere in mainstream reporting is a signal on its own.
- Treat general-purpose AI chatbots as unreliable for this specific task. Grok and Copilot disagreeing on the same photo isn't a fluke, it reflects that these models weren't trained as forensic classifiers. Use dedicated detection tools instead.
The bigger pattern this points to
What happened with McConnell's hospital photos is a preview of something that's going to keep happening around any newsworthy event involving a real person: an information gap, a fabricated image that fills it, a detection tool that catches it, and then a second wave of suspicion that gets pointed at the real evidence once it finally shows up. The watermarking technology worked exactly as designed here. The harder problem, that people now distrust real photos as readily as fake ones, isn't something any watermark can fix. That part depends on rebuilding trust in verification processes themselves, and on institutions being faster and more transparent than a three-week information vacuum, which is what created room for the fake in the first place.
For anyone building products in this space, and there's a real opportunity here, the technical bar isn't just "detect AI-generated images." It's making that detection cross-platform, resilient to screenshots and re-compression, and understandable enough that a Reddit user can trust the result without needing to know what latent space diffusion is.
