Has Google’s AI watermarking system been reverse-engineered?

A software developer, using the pseudonym Aloshdenny, has published a detailed claim of reverse-engineering Google DeepMind's SynthID watermarking system. SynthID is a near-invisible digital watermark embedded into images generated by Google's AI tools, including Gemini Nano and Veo 3, designed to persist even after edits. Aloshdenny's method involved generating 200 pure black images using Gemini, then using signal processing techniques—enhancing contrast, denoising, and averaging patterns—to isolate the watermark's unique frequency signature across color channels. The developer open-sourced their findings on GitHub, noting the process required no neural networks or proprietary access, just significant analysis of the image data.

Despite mapping the watermark's insertion method, Aloshdenny admits they could not fully delete it; the best result was confusing the decoder enough to give a false negative. This partial success led them to praise SynthID's engineering, stating its goal is to 'raise the cost of misuse' rather than be unbreakable. Google has firmly contested the claim. A spokesperson stated it is 'incorrect to say this tool can systematically remove SynthID watermarks,' reaffirming it as a robust tool for AI content identification. The situation highlights the ongoing arms race between content authentication and those seeking to bypass it.