Concerns have been raised that invisible watermarks used to verify the origin of AI-generated content could be misused as a means to track users. Critics say that if identification signals hidden in content are linked to personal data, they could turn into a surveillance tool without users' knowledge.

Gigazine, a Japanese media outlet, reported on Nov. 8 that Brandon Thomas (브랜던 토머스), founder of image and video production software company ArtCraft, argued in a recent article that hidden marks used to track content without user consent should be classified as “Spymark.”

Thomas cited Google’s SynthID as a representative example. The technology inserts digital signals that are difficult for people to notice into AI-generated images, videos, audio and text to verify whether they were generated. Unlike typical metadata, it is designed to leave traces even after image editing or compression.

Thomas focused on the image watermarking technology “SynthID-O” described in a Google research paper, noting it can carry 136 bits of information in a 512 by 512-pixel image. He said 64 bits could be used as an identifier, with the remaining 72 bits used for error correction.

He argued that if the identifier is linked to a separate database containing names, IP addresses, dates of birth and political affiliations, it could be misused to track content creators or distribution routes.

This concern, however, is based on technical possibilities. No evidence was presented that Google actually inserts personal identifiers or tracks users. Google describes SynthID as a technology to verify the origin of AI-generated content.

Thomas also pointed to differences from typical metadata. He said image EXIF data or ID3 information in audio files can be checked, modified or deleted, but invisible marks embedded in the content itself are difficult for users to notice or remove.

Thomas said if such technology expands across social media, smartphones and content creation tools, the possibility of tracking how posts spread increases. He warned in particular of the risk of exposing the identities of whistleblowers or people who express opinions anonymously.

As tools to verify the authenticity of AI-generated material spread, how to balance content transparency with protection of user privacy is emerging as a new challenge.