Google Bakes AI Image Watermark into Pixels

Google has introduced a new feature for its Gemini Nano Banana models that allows users to skip the visual watermark on AI-generated pictures. The feature, which is still rolling out, can be found in the Settings menu as a Media Watermark option. This toggle switch turns off the watermark globally and prevents it from appearing on all images generated using Gemini until it's turned back on.

The visible watermarks on AI-generated images were initially intended to identify them as such but have proven unreliable. They are easily removable, even with basic image editing software like Photoshop, which can strip out the watermark in just 30 seconds. Additionally, a cottage industry of reverse alpha-blending tools has emerged specifically for removing Gemini watermarks.

Google's new watermark is not visible to the naked eye but is instead baked into the pixels themselves. Dubbed SynthID, this system was announced by Google DeepMind in 2023 and is designed to prevent AI-generated images from being mistaken for real ones. While it's not indestructible, researchers have found that it requires significant effort to degrade the watermark, with cropping an image down to about 10% of its original size being one such method.