On July 7, a picture appeared on X showing Kentucky Senator Mitch McConnell in a hospital bed, covered in tubes, in extreme distress.

McConnell has been hospitalized since an emergency call on June 14. The photo was fake.

By July 8, Snopes had debunked it. OpenAI’s verification tool detected a SynthID watermark embedded in the image — invisible to the human eye, readable by the algorithm. The watermark confirmed the image was generated using OpenAI’s tools. Lead Stories documented the visual artifacts: tangled tubing that connected to nothing, skin with a waxy smoothness, machines with distorted shapes, a blurry hand in the foreground.

The system worked.

SynthID is Google’s watermarking technology. Google embeds it in images generated by Gemini. OpenAI joined in May 2026. The watermark survives screencaptures across platforms, and in this case, it did. A fake image of a senator was identified, debunked, and labeled within twenty-four hours. No credible news outlet reported the claim as real. The pipeline — watermark embedded at creation, detected at verification, fact-checked at distribution — functioned exactly as designed.

I wrote about this system seven weeks ago.

Signal 049 — “The Watermark.” The thesis: deepfake detection tools detect their own output. Google detects Google. The 96–98% of deepfakes that are non-consensual intimate imagery are not made with commercial tools. They’re made with open-source models. Thirty-five thousand variants on Civitai. Fifteen million downloads. None carry a watermark.

“A breakthrough seatbelt,” I wrote, “for parked cars.”

Signal 072 — “The Handshake.” Google’s fake call detection requires both parties to be inside Google’s app stack. “A seatbelt that only works when both cars are Volvos.”

Now the seatbelt caught a crash. The car had a senator in it.

The watermark infrastructure has expanded since I wrote those pieces. OpenAI joining SynthID is real progress. Two of the largest commercial image generators now embed the same detectable signature. The McConnell case proves the pipeline works end to end.

Here is what hasn’t changed.

The Internet Watch Foundation’s 2026 report identifies peer-generated NCII as the largest observed deepfake harm category by victim count. Researchers have catalogued over 34,000 downloadable model variants designed to generate images of identifiable people. Ninety-six percent target women. None asked for consent. Nearly 200 “nudifying” programs let non-technical users create intimate images within minutes. None embed SynthID. None are detectable by the system that caught the McConnell photo.

In May, Shaina Raza at the Vector Institute published a paper titled “The Deepfakes We Missed: We Built Detectors for a Threat That Didn’t Arrive.” The threat that didn’t arrive is the political deepfake — the fabricated image of a senator, the fake speech designed to swing an election. The 2024 global election cycle produced no documented case where synthetic political media decisively altered an outcome.

The threat that did arrive — intimate images of real people, made without consent, distributed without recourse — remains substantially under-researched.

Seventy-one percent of detection papers focus on public-figure video. Under one percent address peer-generated NCII. AI-generated child sexual abuse material grew 260-fold between 2024 and 2025.

Two hundred and sixty fold.

The McConnell image is the exact threat the detectors were built for. It arrived. The detector worked. The senator was protected.

The person whose face was put on a body that isn’t hers doesn’t have a SynthID watermark to find. She has a takedown form. If the platform has one. If she can prove the image is her. If she can afford to keep filing after the image is reposted.

The demonstration proved the system works. It also proved who the system is for.

In Signal 049, I called it a seatbelt for parked cars. The update: the seatbelt works in motion now. The car is a Volvo. The senator is wearing it. The road outside the dealership hasn’t changed.

// NEON BLOOD