Google’s SynthID Watermark Debunks First High-Profile Deepfake
Google’s SynthID watermark survived screenshotting across Reddit and X, debunking a viral hoax image of Mitch McConnell in hospital. Snopes confirmed the AI-generated nature of the image. Anthropic does not participate in the SynthID program.
The Invisible Watermark
In July 2026, Google’s SynthID watermarking system achieved its first significant real-world success after Snopes used it to expose a manipulated image of Senator Mitch McConnell. The watermark, invisibly embedded within AI-generated images, resisted screenshotting and sharing across platforms.
How It Works
Launched at Google I/O in 2025, SynthID integrates an invisible signature into AI-generated visuals. This signature is encoded within the image data, ensuring its persistence through resizing, compression, or platform switching. In the case of McConnell’s image, despite being screenshotted and shared widely, the watermark remained detectable.
OpenAI joined the SynthID program in May 2026, integrating the watermark into its image outputs alongside C2PA metadata. However, not all AI tools participate; neither does Anthropic.
Limitations and Challenges
The key constraint of SynthID is its reliance on participating image-generation tools. Gemini models have incorporated the watermark since their launch. Users can verify images by asking a Gemini model or using OpenAI’s public verification tool. Yet, unwatermarked images are produced by open-source models and labs that do not participate in the program.
YouTube has independently introduced automatic labeling for AI-generated videos but faces a broader challenge: watermarking requires widespread adoption, currently voluntary among content creators.
By
Ana Maria Constantin
(Expertise: Digital Marketing, Product Management, Branding & Identity)