Watermark removed, still detected as AI. Why?

Four layers, not one

People think “the watermark” is one thing. An AI image usually carries four, and they behave very differently.

Layer	Can you see it	Survives a crop	Survives re-save	Can it be removed
Corner sparkle	Yes	No	Yes	Yes
SynthID	No	Mostly	Mostly	No
C2PA credentials	No	Yes	Often lost	Yes, very easily
Model fingerprint	No	Partly	Partly	Not really

Removing the sparkle clears row one. Rows two, three and four are untouched. Whichever of them the checker looked at will give you the same answer as before.

Layer 2: the signal inside the pixels

SynthID is not drawn anywhere. It is a pattern of tiny changes spread across the whole image. Each change is far too small to see. They are arranged so a matching detector can measure them.

There is no area to rebuild. That is why no browser tool removes it, including ours.

It is also built to survive what people do to pictures. Saving again, resizing, screenshots, colour grading, filters. Not forever — enough damage kills it — but by then the picture is not worth keeping.

Layer 3: the paperwork

Content credentials are a signed note attached to the file. What made it, when, and what happened to it after. Unlike SynthID, this one is easy to remove. Most editors drop it when you save. Most social apps strip it on upload without asking.

So missing credentials prove nothing. Present credentials are strong evidence. If a checker read credentials on your file, they were still attached.

Layer 4: how the image was made

This one nobody added on purpose. AI models leave small statistical traces in the noise and in the fine detail. Real camera noise and model noise do not look the same under analysis. The “AI detector” websites are trained to spot exactly this.

They are also the weakest layer by far. They are wrong in both directions, and compression or editing makes them much worse. If a percentage from one of those sites flagged your image, that is a guess, not a measurement.

So which one caught you?

• Gemini said it was made with Google AI → SynthID. That is a real measurement, and it will keep saying so.
• A platform showed an AI label on your post → almost certainly content credentials, read at upload time.
• A website gave you a percentage → a classifier. Treat it as a guess.

What it is actually good for

Quite a lot, once you stop expecting it to hide anything. A corner sparkle fights your layout in a mockup. It lands in the bleed on a print job. It jumps in a video edit where every other shot is clean. It skews a dataset or a visual test. These are all real reasons, and none of them need anyone to be fooled.

If the goal is to pass AI media off as a real photo, this site cannot help. The thing standing in the way is not the logo. And in many countries, saying that something is AI-made is now the law.

Questions people also ask

I removed the watermark but Gemini still says the image is AI-made. Why?

Because it is reading SynthID. That is spread through every pixel instead of being drawn in the corner, so removing the logo does not touch it.

Will a screenshot remove the AI watermark?

No. The sparkle is in the pixels, so it is in the screenshot too. And SynthID is built to survive exactly this kind of re-capture.

Does converting to JPG or PNG help?

No. Changing format just saves the same pixels again. It may drop metadata, including content credentials, but the visible mark and the hidden signal both come through.