Someone sends you a picture. You suspect it was generated, but it arrived as a screenshot, so the file itself has nothing in it: no camera fields, no generator record, no signed manifest. That is the most common situation there is, and until recently the honest answer was that nothing more could be done.
SynthID is the exception. It is an invisible watermark Google writes into the pixels of images from Gemini and Imagen, and because it lives in the pixels rather than in the metadata, it does not care that the labels were stripped. Until late last year only Google could read it. Now the check sits in the Gemini app, so anyone can run it.
We wanted to know whether it actually survives being screenshotted, because that is the case people keep bringing us. Google documents the watermark as robust to cropping, filters and lossy compression. It says nothing about screenshots, and a screenshot is harsher than any of those: the operating system redraws the picture at whatever size it happens to be on your screen.
So we tried it. On 12 August 2026, in about ten minutes.
What we did
We generated a portrait in the Gemini app, a man in a wide-brimmed hat on a mountain road, at 1024 by 1024. Then we took a screenshot of it on screen and handed that screenshot straight back to the Gemini app with one question:
Was this image created with Google AI? Check it for a SynthID watermark and say plainly whether one was detected.
The first screenshot had the picture sitting in the middle of the screen with dark margins on either side, which is what a screenshot normally looks like. The answer came back:
No SynthID watermark was detected in this image, which indicates it was not created using Google AI generation models.
That is a confident all-clear on an image Gemini itself had made two minutes earlier. If you were checking a photograph somebody sent you, that answer would have settled the question the wrong way.
Gemini did volunteer why, in the next paragraph:
Also, since this image has large black borders around it, cropping the image as close as possible to the main subject before verifying can help yield the most accurate result.
So we took a second screenshot of the same picture, this time filling the frame edge to edge with no margins, and asked the same question. Same image, same session, same wording:
All or part of this image was created or edited with Google AI, as a SynthID watermark was detected.
What that means for you
The mark does survive a screenshot. What breaks the check is everything around the picture: the desktop, the browser chrome, the black bars, the message bubble it was sitting in.
That makes sense once you think about how the watermark is read. The detector is looking for a statistical pattern spread across the image. Pad that image with a large area that carries no pattern at all and you have diluted the signal, exactly as you would by shrinking the picture into a corner of a much larger canvas.
The practical instruction is one line: crop tight to the picture before you check it. Any photo editor will do, including the one already on your phone. Cut away the wallpaper, the toolbar and the borders until the frame contains the image and nothing else.
We have not seen this written down anywhere, which is the reason for this post.
The limits, honestly
This only answers for Google's own tools. SynthID is in Gemini and Imagen output, so a picture from Midjourney, Ideogram or Stable Diffusion will come back clean whether or not it was generated. A negative result is not evidence of anything.
It also means giving Google the file. That is a real cost, and it is your decision rather than ours. SynthCheck does not do it for you and never uploads anything: the check we run happens entirely in your browser and reads the file, not the pixels. When a file has nothing left to read, we say so and point you here, because at that point the watermark is the only thing left that could answer.
And one caution about the answer itself. "Created or edited with Google AI" covers a real photograph that was retouched with a Google tool. It is not the same claim as "this picture is fake", and it should not be reported as one.
The short version
If a suspicious picture reached you as a screenshot, crop it tight to the image, then open the Gemini app and ask whether it was made with Google AI. Uncropped, the same file gave us a confident wrong answer. Cropped, it gave the right one.
Check the file first, here in your browser, before you hand it to anybody. Sometimes the answer is still sitting inside it.