Google has expanded its SynthID watermarking system to the Gemini app, allowing users to check whether an image was produced by Google’s AI models. The feature is available on mobile starting today. Users can upload an image through the Gemini interface and type a query such as “is this image AI generated?” The app will then scan the invisible watermark and provide context about its findings. Google plans to extend the SynthID system to audio and video in future updates.
SynthID was introduced in 2023, and Google reports that more than twenty billion pieces of AI generated content have since been tagged with it. In addition to the invisible markers, Google will keep a visible Gemini watermark on images created by users of the free and Pro tiers. Images generated by Ultra tier subscribers will not carry the visible symbol to maintain a cleaner appearance for professional work.
Uploading an image for verification is simple. Users select the plus icon in the Gemini app, choose a file from their device and ask whether the image was generated by AI. If SynthID is present, Gemini will confirm it. If not, Gemini uses its own analysis to provide a best estimate. The accuracy of this secondary judgement depends on visual clues rather than embedded markers.
Credit: GoogleTesting shows the limits of the system. Images generated by Google AI are reliably identified, with both the invisible and visible watermarks detected. Images created using platforms without SynthID, such as ChatGPT, produce less certain results. Gemini can indicate that no Google watermark is present but can only give an opinion based on the image’s characteristics. When tested with unrelated images found online, Gemini responded that the image was not created with Google AI but could not determine whether it was made with other tools.
Google’s approach moves toward clearer AI origin detection, but the broader landscape remains fragmented. Without a unified watermarking standard across major AI image generators, tools like SynthID provide accurate results only within their own ecosystem. Until wider adoption occurs, AI image verification will continue to depend on a mix of embedded markers and probabilistic reasoning.
