Tested: Google SynthID works great, but labeling AI content may be a losing game - Ars Technica

What the report says

Ars Technica reported that Google’s SynthID watermarking system performed strongly in its own hands-on test, even as the article cautioned that labeling AI-generated media remains a difficult long-term challenge. The report frames the issue against the rapid growth of synthetic media, citing Starling Lab’s estimate that generative AI produced 1.5 billion images in about 18 months and Google’s statement that its tools have generated more than 100 billion AI images and videos in recent years.

The article explains that Google uses both SynthID and C2PA-style provenance metadata to identify AI content. C2PA can be cryptographically verified but is easy to remove through routine edits or screenshots. SynthID, by contrast, is embedded into the media itself. Google DeepMind scientist Pushmeet Kohli told Ars the system was built with expected attacks and transformations in mind, including cropping, filters and other edits.

In Ars’ test, images created or edited with Google’s Nano Banana Pro were repeatedly compressed and resized using a Python script, with cropped versions checked along the way. SynthID remained detectable after hundreds of simulated sharing cycles and even after screenshots. The watermark eventually failed when heavily degraded images were cropped substantially, including a 20 percent crop after 300 compression rounds.

Ars concluded that SynthID appears more durable than simple metadata and some rival watermarking efforts, but not definitive. Google does not claim it can withstand every adversarial attack, and wider adoption by companies such as OpenAI, Runway and Nvidia could make it a larger target. The broader significance is that watermarking may help identify some AI media, but it is unlikely to settle questions of online authenticity by itself.