A software developer known as Aloshdenny claims to have reverse-engineered Google DeepMind’s SynthID watermarking system, revealing a method to manipulate the invisible digital tags embedded in AI-generated content.
Posting his findings on GitHub and Medium, the developer documented a process that uses signal processing to identify and interfere with the watermark's frequency patterns. The method does not require proprietary access or neural networks to function.
Aloshdenny explained that the breakthrough came from analyzing 200 Gemini-generated images that were essentially pure black. "Turns out if you’re unemployed and average enough ‘pure black’ AI-generated images, every nonzero pixel is literally just the watermark staring back at you," he wrote on Medium.
Disrupting the decoder
The developer's process involves enhancing the contrast and saturation of these black images to expose the underlying watermark patterns. By averaging these patterns, he identified the magnitude and phase of the watermark signal across various frequency bins.
While the developer claims he can target these frequencies to confuse SynthID decoders, he noted that he has not successfully deleted the watermark entirely. Instead, the technique aims to disrupt the systems designed to read the watermarked pixels.
"The fact that the best I could pull off was confuse the decoder enough that it gives up — not actually delete the thing — says a lot about how well it was designed," Aloshdenny said. He noted that the system's goal is likely to increase the cost of misuse rather than provide absolute security.
Google has dismissed the claims of a systematic breach. The company maintains that the watermarking technology remains effective at identifying AI-generated media.
"It is incorrect to say this tool can systematically remove SynthID watermarks," Google spokesperson Myriam Khan told The Verge. "SynthID is a robust, effective watermarking tool for AI-generated content."
SynthID is currently integrated into several Google products, including the Veo 3 video model and Nano Banana. The system embeds metadata directly into the pixels of an image at the moment of creation to ensure traceability.
