This Is Not a General AI Detector

Google has opened its watermark detector to everyone. Upload an image, a video or a clip of audio and it will tell you whether the file carries Google's invisible mark. The catch is printed on the tool itself: it is not a general AI detector.

On Tuesday, Google switched on synthid.com for the world, in English, for anyone with an account. Since 2023 it has been hiding SynthID, an invisible watermark, inside media made by its own models. Now you can bring a file to the portal and get a straight answer: was this made, or altered, by Google or one of its partners?

The partner list is the real story. It is not just Google. OpenAI, NVIDIA and Kakao sign their AI media with the same invisible ink, and Apple is next. For the first time there is a single place to check watermarks from several of the largest labs at once.

What it can actually see

Google says it has watermarked more than 180 billion images and videos, and 240,000 years of audio. The detector gives a plain yes or no. The same check already runs inside Search, the Gemini app and Chrome, where Google says it handles about a million requests a day.

I mapped the guest list

I downloaded Google's announcement and the detector's own notes, then mapped what it can see against the tools people actually reach for. The covered list is short. Google's own models lead it: Gemini and Nano Banana for images, Veo for video, Lyria for audio. Then OpenAI's images, NVIDIA's Cosmos video, and Kakao, with Apple next. Almost everything else is invisible to it, because those makers never joined the club.

You can see the shape of it below. Coverage is not a switch; it is a list you assemble, lab by lab, and it only ever covers the names on it.

The hole in the middle

Here is the sentence the tool puts on itself: "This is not a general AI detector." It only knows media from companies that signed up. A Midjourney render, a Stable Diffusion still, or the output of a model running on your own laptop sails straight through with no mark, not because it is authentic but because its maker is not on the list. If you read a blank result as a clean bill of health, the tool has done the opposite of its job.

A watermark you can trust is useful. The photo that carries nothing is the one to worry about.

The honest counterpoint

Google is careful to say all of this, and the design is honest. SynthID is a provenance signal, not a lie detector, and getting rivals to share one portal is genuine progress. The marks are built to survive cropping, filters, dropped frames and compression. Text is the odd one out. Google watermarks Gemini text separately and leaves it out of this detector, and OpenAI's own testing found that swapping a quarter of the words in a passage drops detection from about nine in ten to under two in ten. Words are the easiest thing to fake and the hardest thing to tag.

What would change my mind: an independent test showing the marks survive a round trip through a social network, and a detector willing to say "unknown" instead of staying silent.

If you are working out which parts of your stack you can actually trust, email brandon@kreostudio.co.uk.