SynthID: How Google Detects AI Photo and Video Deepfakes

Google's SynthID technology helped Snopes debunk an AI-generated photo of U.S. Senator Mitch McConnell in July 2026. It is the first high-profile case in which an AI-content detection system actually stopped a political deepfake from spreading on major platforms.

The image that spread across Reddit and X

The image appeared on Reddit and X and quickly collected thousands of reposts. It showed the 84-year-old Republican senator from Kentucky on a hospital bed, tangled in tubes and in apparent extreme distress. The picture looked convincing, and users shared it as real news about the politician's health without questioning its authenticity.

Snopes ran the image through the SynthID detector and got an unambiguous answer: it was an AI deepfake, not a photograph. The newsroom described the outcome as the first "high-profile real-world win" for Google's system in the fight against AI-generated disinformation.

How SynthID works

Google DeepMind launched SynthID in 2023 as an invisible watermarking system for AI-generated content. When a model creates an image, it embeds a unique cryptographic pattern in the pixels. The human eye cannot see it, but a specialized verifier can read it unambiguously. The watermark survives file compression and resizing.

Key characteristics of the technology:

• Developed by Google DeepMind, launched in 2023
• Supports images, text, audio and video
• Survives file compression and resizing
• An open verifier is publicly available

There is a fundamental limitation: SynthID only detects content created by generators that support the technology. A deepfake made with third-party tools that lack SynthID will simply go undetected.

Why this precedent matters beyond a single story

Since its announcement, SynthID had remained largely a technical promise: Google built it into its own generators and showed it off at conferences, but no publicly documented cases of real use in editorial work had been recorded until now.

The McConnell case changed that. An independent fact-checking service applied the tool on its own against a political deepfake with mass reach and got a verifiable result. It is the first time watermarking technology has actually stopped the spread of a specific fake about a living politician, and it happened in an organic editorial process, without Google's direct involvement.

As AI-generated fakes about politicians multiply, a case like this will be a weighty argument in the debate over mandatory labeling of AI content, a debate already under way in the EU, the U.S. and other jurisdictions.

What this means

SynthID has passed a field test: AI-content watermarking has stopped being an abstraction and become a working editorial tool. The documented precedent will speed up its adoption among AI image providers and push forward the legislative debate on industry labeling standards.

Frequently asked questions

Can SynthID detect any deepfake?

No. The system only detects content created by generators with built-in SynthID support. If a deepfake was made with a third-party tool without this technology, there is no watermark and nothing to detect.

Where can I check an image for a SynthID watermark?

Google DeepMind offers an open SynthID verifier available to the general public. A check takes a few seconds and requires no special technical knowledge.

How does Google recognize AI photos and video?

Google uses SynthID, invisible watermarks embedded in AI-generated images. Even after the photo is processed, the system recognizes it as AI-made.

How did this help expose the senator deepfake?

Snopes detected the SynthID watermark on the photo of Senator McConnell circulating on Reddit and X and proved its AI origin. It is the first high-profile case of a detection system stopping a political deepfake.