Google’s SynthID AI Watermarking Tech Claimed to Be Reverse-Engineered

Google’s synth ID is a distinct exception to artificial intelligence (AI) detection in the world of artificial Intelligence (Ai) discovery. Originally first seen in 2023, it is watermarking technology that adds an invisible and imperceptible layer of water mark to text images video audio (including text), photos. The tech giant, based in Mountain View, says SynthID cannot be removed but one developer is now saying that he was not only able to reverse engineer the watermarking tech (but also distorted it to the level that it confuses systems detecting AI content).

Can Google’s SynthID Be Removed?

There is no short answer to the question “what’s a “? at least not completely – but to the extent that . But even if developer Aloshdenny claims are to be believed, the watermarking technology is not the impenetrable wall Google says it is. Note that the whole process was not ‘breaking into Google’s watermarking models or manipulating encoders/decoder,’ interestingly noted here.

In a Medium post, developer explained how the pattern of SynthID was detected by 100 plain white and 100 pure black images created by the tech giant’s Nano Banana model. He also used 1,23,268 pairs of original and AI-edited images with the same model alongside.

The developer then added the contrast of the black and white images, before increasing saturation and denoising them. It was conceived that on a pure black Nano Banana image, every non-zero pixel value should represent the watermark. But observing it alone is not enough, since Google’s encoder uses an algorithm that has a randomness factor (i.e., every image may have arguably unique position of the SynthID).

After the full-time reverse engineering, the developer then tried to remove the watermark. But he said ‘It was very hard to understand, and I did not think Google had implemented it so solid that we could remove the pattern without distorting the image. But he does come up with a trick to confuse the decoder so that it cannot know whether an image has the watermark or is not confident enough.

Aloshdenny has publicly released his work on GitHub so that others can review the process and verify it is true. A Google spokesman told The Verge that the claims that SynthID is removable are not true, and “the technology remains an effective method to label and detect content generated by Gemini and other Google AI models.”