Spymarks, Not Watermarks

A new type of "watermark" called a "spymark" is emerging, designed to trace and identify digital content without the creator's consent. While traditional watermarks are visible marks used to verify authenticity or claim ownership, spymarks are covert signals that make content traceable. Google's SynthID is an example of a spymark, embedding hidden signals into images, audio, text, and video that can encode database identifiers linked to users' identities.

These identifiers include personal information such as user records, full names, IP addresses, date of birth, physical addresses, political party affiliations, and more. Google's SynthID-O variant can encode a 136-bit payload in a 512x512-pixel image, leaving 72 bits for error correction. Other tech companies, including OpenAI, are also developing spymark systems that claim to help identify AI-generated content.

Spymarks can operate in various media, including audio, where they are typically inaudible but can encode personal information using similar principles to image spymarks. These hidden signals can survive compression, re-encoding, and metadata removal, making them difficult to remove or edit. The term "watermark" is being repurposed to encompass these new spymark technologies, emphasizing the privacy concerns they present.