Developer claims Google SynthID watermark can be copied or weakened

In short: A developer says they figured out parts of Google’s SynthID watermark for AI images, but Google disputes the claim.

What happened

A software developer using the name Aloshdenny published a write-up and open-source code that they say reveals how Google DeepMind’s SynthID works. SynthID is a near-invisible watermark, meaning a hidden label inside an image, that marks it as made by Google’s AI tools.

Aloshdenny claims their method can weaken SynthID in some AI-made images so that watermark readers, called detectors, struggle to recognize it. They also claim a watermark could be inserted into other images. The developer says they did not use special access, and instead generated about 200 test images using Google’s Gemini and then analyzed the patterns.

In simple terms, the approach is like making lots of blank test pages, then turning up the contrast to see the faint “stamp” that keeps showing up. By averaging many images together, the watermark pattern becomes easier to spot, and that can help someone try to hide it or mimic it.

Google strongly disagrees with the idea that SynthID has been reliably cracked. A Google spokesperson told The Verge, “It is incorrect to say this tool can systematically remove SynthID watermarks,” and said SynthID remains “robust” and “effective.”

Why it matters

Watermarks like SynthID are meant to help people tell what content was made by AI, especially as realistic fake images spread online. If watermarks can be easily removed or copied, it gets harder for regular people, journalists, and platforms to trust quick “AI or not” checks.

Source: The Verge AI