Gemini and Samsung Galaxy AI Watermarks: The Marks on Your Phone Photos

You used Galaxy AI to erase a stranger from a holiday photo. You did not generate anything. And there it is in the corner of the saved image: a small mark announcing AI.

It feels disproportionate — and it reveals how the rules actually work. The moment software invents pixels that a camera sensor never captured, your photo is partly synthetic, and the marking machinery treats it accordingly.

This is a spoke of our AI image watermarks pillar guide.

What lands on the file

The pattern repeats across every generator in this hub: the visible layer and the metadata layer are yours; the invisible layer is not.

Why an edit counts as generation

Traditional editing rearranges what the sensor recorded — exposure, crop, colour. Generative editing fabricates content that was never there: filling a removed object, extending a frame beyond its borders, replacing a sky.

That distinction is what the industry's transparency commitments and the emerging disclosure laws are built on. Provenance schemes are explicitly designed to record *how* a file came to look the way it does, including edit steps, which is why a single generative erase is enough to earn a tag. The legal shape of this is covered in AI image labelling laws and agreements.

Removing the corner mark

A corner mark occupies a small, predictable region:

• Crop — trims the marked edge. Cleanest output, small framing cost.
• Patch — rebuilds the region from surrounding pixels. Keeps the full frame; near-invisible over sky, wall, water or blur, harder over fine texture.

Removing the tags

This is the part most people never look at, and it is where the real identifying information sits: generator names, model versions, edit-provenance entries, source-type declarations, sometimes a signed manifest, plus the ordinary phone EXIF — which on a phone photo can include the location and time the original was taken.

Re-encoding from raw pixels produces an output file with none of it. Not redacted, not overwritten: the container simply is not there.

SynthID: the part we cannot promise

Gemini imagery is the most likely place a consumer will encounter genuinely robust invisible watermarking. SynthID embeds a statistical pattern across the pixel values during generation, redundantly, and is tested against compression, resizing, cropping and screen capture.

Our tool offers optional resample and dither strengths that can degrade fragile pixel-domain schemes. We label it best-effort because it is, and it visibly costs you quality. For the full engineering reasoning, read our honest SynthID guide.

The routine

1. 1Get the photo onto any device with a browser.
2. 2Crop or patch the visible corner mark.
3. 3Re-encode to empty the metadata container.
4. 4Read the before-and-after report and see exactly what was found and removed.

Read next: what each watermark type actually survives →