Google's invisible watermark survives edits, but won't stop AI misinformation

The volume of synthetic media entering circulation each day rivals what took the entire history of photography to accumulate. A research collaboration between Stanford and USC found that from the camera's invention through 1975, humanity created 1.5 billion images—a milestone that generative AI systems reached in just 18 months.

Google revealed at its spring I/O conference that its platforms alone have generated more than 100 billion AI images and videos in recent years. Now the company is betting that invisible watermarking can help address the authenticity challenge posed by this explosion of machine-made content. Google is rolling out SynthID across partnerships with major players including OpenAI, Runway, and Nvidia.

How watermarking differs from metadata#

Two main strategies exist for flagging synthetic content. Invisible watermarks like SynthID embed data directly into pixel information or audio waveforms, while metadata schemes such as the Coalition for Content Provenance and Authenticity (C2PA) store verification information separately.

The critical difference is persistence. C2PA is cryptographically secure against forgery but vanishes easily—a simple screenshot, crop, or export removes it entirely. In contrast, SynthID persists through the wear and tear that happens as content circulates online: compression, resizing, and edits. Google DeepMind researcher Pushmeet Kohli emphasized to Ars Technica that the team invested significantly in building SynthID's durability, with the goal of making the watermark detectable even in widely-shared, heavily-modified content.

The labeling problem#

While SynthID's technical resilience is impressive, the larger challenge remains unsolved: knowing that something is AI-generated doesn't automatically prevent its misuse. Watermarking addresses identification but not the underlying incentive to spread misleading content. The flood of synthetic media—and the platforms distributing it—suggests that labeling alone may not be a sufficient tool against AI-driven misinformation.