Visible AI watermarks: what they show and what they do not
Some AI tools and platforms put a visible label on the picture itself. These marks are the easiest signal to see and the easiest to remove. A visible mark shows that a tool or platform added it, or that someone drew it there. A missing mark shows nothing. This guide lists the marks we can source and what each one is worth.
The New Watermark image checker does not look for visible marks. You look, and the checker reads the signals in the file. Visible marks are one of five kinds of signal; How to tell if an image is AI compares them all.
The marks we can source
| Mark | Where it comes from | What to know |
|---|---|---|
| Gemini sparkle | The Gemini app | Reported: since August 14, 2026 it is optional, through Settings and then Media Watermark. Turning it off is reported to leave SynthID and Content Credentials in place. |
| "Imagined with AI" label | Meta | Described in Meta's February 2024 announcement as the label for images made with Meta AI, alongside invisible watermarks and IPTC metadata. |
| "AI-generated content" text | Samsung Galaxy AI | Reported to appear at the bottom left of images. Samsung is also reported to add Content Credentials to AI-edited images. |
| Coloured squares | DALL-E 2 | Reported, and we rate the sourcing as weak. A legacy mark from an older model. |
| Small "b" | Microsoft's Bing | Reported in 2023, by secondary sources only. |
Many tools add no visible mark that we could find documented, including Midjourney, Ideogram, Leonardo and Canva. The Stable Diffusion reference scripts, Automatic1111 and ComfyUI add none. We found no documented visible mark for OpenAI's gpt-image models, Adobe Firefly or Google's Imagen.
Law can also drive the choice. It is reported that China's standard GB 45438-2025, in effect since September 1, 2025, requires an explicit visible label on AI-generated content, along with implicit labels in metadata. That is one reason some tools add a visible mark and others do not.
What a visible mark proves
It proves very little. A visible mark is pixels, and pixels can be edited by anyone.
- It can be removed. Cropping, painting over or retouching can take a mark away. Some tools let people switch it off, and the Gemini app is reported to be one of them.
- It can be added. Anyone can place a label on a photograph that has nothing to do with AI. A mark is not a signature.
- It tells you where to look. A mark naming a company is a good reason to run that company's checker on the file, because the stronger signals may still be there.
No visible mark does not mean an image is real or human-made. It means no one added a label, or someone took it off.
Why the checker never detects marks automatically
We could try to match corner patterns, but that is brittle, and the cost of being wrong falls on the picture. A wrong "mark found" would label an unrelated photograph. A wrong "no mark" would read to most people as "not AI". The Gemini sparkle is also optional now, so its absence tells you even less than it did. For these reasons the checker leaves visible marks to you. If you can see Meta's "Imagined with AI" label, you can tick a box on the checker and it will add Meta's identification tool to the list of links.
What to do when you see a mark
- Note which mark it is and which company it points to.
- Run the file through the checker to read its signed records and metadata.
- For a Gemini sparkle, also try Google's SynthID Detector. The invisible watermark is reported to stay when the visible mark is off. What is SynthID? explains the detector.
- For Meta's label, you can try Meta's identification tool, reported in July 2026 as a preview for Meta's own watermark on output from its newest image model. Reports say it missed some cropped images.
- Remember that each of those links opens another company's website, may ask you to sign in, and receives any file you upload. We are not affiliated with them.
Frequently asked questions
- Can New Watermark detect the Gemini sparkle?
- No. A false "found" would label a real photograph, and a false "none" would read as "not AI", so we leave visible marks to your own eyes.
- If a picture has no visible mark, is it real?
- No. Marks are optional, removable and absent from many tools.
- Can a visible mark be faked?
- Yes. Anyone can draw a label onto any picture.
- Does removing the visible mark remove the invisible watermark?
- For the Gemini app, it is reported that turning the visible mark off leaves SynthID and Content Credentials. We have no information on what other edits do. Editing can damage watermarks and records.
- Which visible marks are well documented?
- Meta's "Imagined with AI" label is described in Meta's own announcement. The others in the table rest on reports.
- Is a visible label always required?
- Rules differ by country, and we have not surveyed them. China's GB 45438-2025 is reported to require one.