Google's SynthID watermark has successfully exposed its first notable deepfake: a fabricated image of Mitch McConnell.

Google's SynthID watermark has successfully exposed its first notable deepfake: a fabricated image of Mitch McConnell.

      TL;DR: Google’s SynthID watermark helped debunk a viral fabricated image of Senator McConnell in a hospital. The watermark survived screenshotting across various platforms, marking its first major successful application in the real world.

      This week, Google’s SynthID watermarking system achieved a significant milestone when Snopes used it to dismiss an AI-generated image of Senator Mitch McConnell that had gained traction on Reddit and X. The image, which depicted McConnell connected to tubes in a hospital bed and appearing to be in distress, was identified as AI-generated after Snopes discovered the embedded invisible SynthID watermark.

      Since his hospitalization following an emergency call on June 14, there has been rampant speculation regarding McConnell’s health, as he has remained mostly out of the public eye. In this instance, the claims were entirely fabricated, and the watermark functioned precisely as intended.

      SynthID, which was introduced at Google I/O in 2025, embeds an invisible signature within AI-generated images. This signature is part of the image data itself, allowing it to endure screenshots, resizing, and compression on various platforms. This durability was crucial in debunking the McConnell image, as it was shared widely, yet the watermark remained intact. OpenAI joined the SynthID initiative in May 2026, incorporating the watermark with C2PA metadata in its image creations, though Anthropic is not a participant.

      The primary limitation of SynthID is that it only functions when the image-generation tool is actively involved. Since their launch, Gemini models have included the watermark. Users can verify images by querying a Gemini model or uploading them to OpenAI’s public verification tool. However, images generated by tools outside of the initiative, such as open-source models and those from non-participating organizations, do not feature a watermark. Additionally, YouTube has initiated the auto-labeling of AI-generated videos, but the overarching issue remains that watermarking can only be effective with broad adoption, which is still voluntary.

Other articles

AI Financial Update: Five Significant Funding Rounds Today AI Financial Update: Five Significant Funding Rounds Today A $300 million quantum funding round, a billion-dollar valuation for an AI agent, and Europe's investments in voice and energy: highlights from the last 24 hours in AI and deep-tech financing. HubSpot retracts its decision regarding the use of customer data for AI purposes. HubSpot retracts its decision regarding the use of customer data for AI purposes. HubSpot has abandoned its proposal to aggregate customer data for an AI feature following a backlash that lasted four days, highlighting how sensitive the issue of AI and data collection has become. NATO's AI 'Kill Web' along its eastern border NATO's AI 'Kill Web' along its eastern border NATO is developing an AI-driven 'Kill Web' composed of sensors, drones, and satellites positioned along its eastern border to identify and respond to a Russian assault before it occurs. Venus Aerospace secures $91 million for its detonation engine. Venus Aerospace secures $91 million for its detonation engine. Venus Aerospace has secured $91 million in a Series B funding round to develop its rotating detonation rocket engine into equipment for missiles and high-speed space vehicles. Adult content creators targeted by scammers are having government websites taken down. Adult content creators targeted by scammers are having government websites taken down. Fraudsters are exploiting the names of adult creators to infect reputable government and university websites, as copyright claims discreetly drive thousands of harmful pages out of Google's search results. ChatGPT Live could make interacting with AI seem like something from a film. ChatGPT Live could make interacting with AI seem like something from a film. OpenAI has introduced GPT-Live, an innovative voice model for ChatGPT Voice that can simultaneously listen and speak, enhancing the speed and naturalness of AI conversations.

Google's SynthID watermark has successfully exposed its first notable deepfake: a fabricated image of Mitch McConnell.

Snopes employed Google's invisible SynthID watermark to verify the authenticity of a viral AI-generated image depicting Senator McConnell in a hospital bed. The watermark remained intact even in screenshots.