Field Note · 013
The pixels that confess.
On invisible watermarks and the burden of proof. Summer 2026.
The image standard now has two layers doing different jobs. C2PA, the provenance specification backed by Adobe, Microsoft, Google, and OpenAI, attaches a signed manifest describing who made a file and how. SynthID, from Google DeepMind, hides a second mark inside the pixels themselves, engineered to survive a screenshot, a crop, and compression, invisible to the eye and readable only by a paired detector. The manifest is rich and easy to strip. The watermark is durable and says almost nothing. Under the EU AI Act, whose transparency rules arrive in August 2026, some machine readable version of this marking stops being optional.
We spent a century learning to believe our eyes. Now the image arrives with documents, and we are learning, slowly, to check them at the door. The strange part is not that pictures can lie. Pictures always could, with enough darkroom patience. The strange part is that we now need them to swear they are telling the truth, and that we have taught them to do it in a language only another machine can read. Authenticity used to be something you saw. It is becoming something you look up.