Testing finds Google SynthID watermark survives heavy edits but can be cropped out

In short: Google’s SynthID can still flag many heavily edited AI images, but simple cropping can remove the label and many AI images may never be labeled at all.

What's going on

AI-made photos and videos are being created at huge scale. Starling Lab, a research group tied to Stanford and USC, says it took humanity until 1975 to create about 1.5 billion photos, but AI systems matched that in 18 months. Google has also said its tools helped create more than 100 billion AI images and videos in just a couple of years.

To help people tell what is AI-made, Google uses SynthID, an invisible watermark. Think of it like a faint pattern woven into the picture itself, not a visible stamp. Ars Technica tested SynthID by repeatedly compressing and resizing AI images hundreds of times to mimic what happens when files get shared online.

The test found SynthID remained detectable even after 300 rounds of heavy quality loss, and it even survived screenshots. However, after heavy compression, cropping the image by about 20 percent was enough to make SynthID undetectable. Larger crops could break detection earlier.

What to watch

Even if SynthID works well, it is not a complete answer. Access to checking is limited, and different companies may use different detectors that do not recognize each other’s watermarks. More importantly, people can generate AI images with tools that add no label at all, including models that run on a personal computer.

That is why some researchers argue that proving something is real may matter more than proving something is AI. One approach is C2PA, which adds tamper-evident information to files (like a sealed label on a package), but it is easy to remove and it is not widely used yet.

Source: Arstechnica