Two layers of AI provenance
Metadata — EXIF, XMP, IPTC and C2PA — lives in separate blocks of the file. It is easy to read and easy to remove without touching the image. An invisible watermark is different: it is a subtle pattern spread across the pixel values, invisible to the eye but recognizable by a detector.
| Metadata (C2PA, IPTC, EXIF) | Invisible watermark (e.g. SynthID) | |
|---|---|---|
| Where it lives | Separate blocks in the file | Inside the pixel values |
| Who can read it | Any metadata viewer | Only the provider’s detector |
| Survives screenshots and re-saving | Usually not | Designed to |
| Removed by a metadata cleaner | Yes | No |
See also: What are C2PA / Content Credentials?
What SynthID is
SynthID is a watermarking technology developed by Google DeepMind for content generated by Google’s AI models. It embeds an imperceptible signal during generation and is designed to remain detectable after common changes such as cropping, resizing, compression and color adjustments. Google uses the same family of techniques for images, audio, video and text.
How to check an image for SynthID
Unlike metadata, a watermark cannot be read with a generic viewer — you need the provider’s detector. Google has introduced ways to check whether content contains SynthID, such as its SynthID Detector and verification inside its own products. Other providers use their own watermarking and detection methods.
The report on this site shows metadata only. Its “No AI signals” verdict therefore never rules out a pixel watermark.
See also: How to Check If an Image Is AI-Generated (Metadata, C2PA and Beyond)
Which AI images have invisible watermarks?
- Google’s image models, such as those used in Gemini and Imagen, embed SynthID.
- Meta has said images from Meta AI include an invisible watermark alongside metadata.
- Other hosted services may use their own watermarks; policies vary and change.
- Local tools such as Stable Diffusion usually add no watermark by default.
See also: What Metadata Do AI Image Generators Add? DALL·E, Firefly, Midjourney, Gemini and Stable Diffusion
Why metadata tools cannot remove watermarks
A lossless metadata cleaner copies the image data byte for byte — that is precisely what guarantees zero quality loss. A watermark inside the pixels therefore stays exactly where it is. Tools that promise to “remove AI watermarks” either do not work as claimed or alter the image heavily, and attempting to defeat a provider’s watermark may also breach its terms of use.
What this means in practice
Cleaning metadata is the right tool for privacy and for removing prompts, locations and provenance data you do not want to share. It does not make an AI image undetectable. When you publish realistic AI-generated content, follow the disclosure rules of the platform you post to.
See also: Why are real photos labeled “Made with AI” / “AI info” on Facebook and Instagram? · Which Platforms Label AI Images? Instagram, Facebook, TikTok, YouTube and More
Questions about SynthID and AI watermarks
Does this tool remove SynthID?
No. The tool only removes metadata and never modifies pixels, so any invisible watermark remains in the image.
How can I tell whether an image has SynthID?
Only with Google’s own verification tools. The metadata report on this site cannot see watermarks inside the pixels.
Do all AI images have invisible watermarks?
No. Some providers watermark their output, many do not, and images from local tools usually carry no watermark at all.
Does a screenshot remove SynthID?
It is designed not to. SynthID is built to survive common transformations such as screenshots, cropping and compression.
Is SynthID the same as Content Credentials?
No. Content Credentials are signed metadata stored next to the image; SynthID is a signal inside the pixels. Some systems combine both.
Can platforms detect SynthID?
Only where they have access to a compatible detector. Most platform labels today are driven by metadata and creator disclosure.