The mark that goes, the mark that stays
The change is as concrete as a single toggle. A media watermark option is arriving in Gemini's settings menu; switched off, the sparkle in the corner of images, videos and music made with the Nano Banana and Omni models stops appearing. The setting works in Gemini and in Flow, Google's video editor.[1]
Only the layer you can see with your eyes switches off. Invisible SynthID marks and C2PA provenance data stay inside the file either way. Josh Woodward, vice-president of Google Labs, Gemini and AI Studio, says transparency rests on those two markers, and that whether a piece of content carries them can be learned by asking Gemini or Search. The check moves, then: it stops being something you do by looking and becomes something you do by asking.[1]
Who keeps the ability to read the mark?
In the same week Anthropic said it would embed invisible watermarks in the text Claude produces and attach signed provenance data to supported image files, with detection running for now only inside the company. Arthur Goldstuck, managing director of World Wide Worx, says a watermark on its own cannot establish who authored a finished work, how far the AI was used, or whether any rule was broken. Johan Steyn, founder of AIforBusiness.net, points out that converting a document or running it through a humanising tool can strip the marks away.[2]
Seen from the street, the difference is this: the visible mark was the one check that needed no tool at all. An image landing in a group chat either had the sparkle or it did not. Once that layer becomes optional, both remaining markers need an instrument to be read, and the companies that apply the markers are the ones running that instrument today. The opposite reading is available too: the visible mark was already being cropped away, and image generators from OpenAI and Meta never applied one, so the layer now going may have been doing little work in practice. Both readings arrive at the same gap. Checking from outside runs through a door held by whoever applied the mark.[1], [2]
What would change this?
In an earlier column on AI writing detectors, PicoByte argued that there is still no reliable way to verify from outside whether a person or a model wrote a given text. On the image side the situation goes one step further: a readable marker exists there, but the ability to read it has not been shared. So the thing to wait for is a capability: the marker becoming readable outside Google as well. If Google or a third party opens a published way to read SynthID or C2PA markers outside Google's own surfaces by 31 December 2026, the check becomes something a user can run again. If that does not happen, the only route left stays asking Gemini or Search.[1], [3]