Two separate layers
Metadata — EXIF, XMP, IPTC, C2PA — sits in its own blocks inside the file, easy to read and easy to remove without touching the picture. A pixel watermark is another thing altogether: a faint pattern spread across pixel values that the eye cannot see but a detector can.
About SynthID
SynthID is Google DeepMind’s watermarking technology for content generated by Google’s AI models. It embeds an imperceptible signal during generation and is designed to stay detectable after common changes such as cropping, resizing, compression and color tweaks. Related techniques cover images, audio, video and text.
Detecting a pixel watermark
A generic viewer cannot read a watermark; you need the provider’s own detector. Google has introduced ways to check content for SynthID, including verification inside its products. Other providers have their own methods.
Why washing leaves it alone
A lossless metadata tool — AI Label Remover included — copies image data byte for byte; that is exactly how it guarantees no quality loss. Any watermark in those bytes stays put. Tools that promise to “strip AI watermarks” either do not deliver or alter the image heavily, and trying to defeat a provider’s watermark may breach its terms.
In practice
Washing metadata with the AI metadata remover is the right move for privacy and for removing prompts, locations and provenance you do not want to share. It does not make an AI image undetectable. When posting realistic AI content, follow the platform’s disclosure rules.
Metadata, watermark or fingerprint?
Three different techniques get called “AI labels”, and they behave very differently:
- Metadata — EXIF, XMP, IPTC and C2PA blocks. Anyone can read them, and they are easy to remove or lose when a file is re-saved.
- Pixel watermark — a hidden pattern in the image itself, like SynthID. Only the provider’s detector can read it, and it is built to survive edits.
- Fingerprint — a compact description of the image kept on a server. A lookalike image can be matched against it later, even with no metadata at all.
Who uses invisible watermarks
Google applies SynthID to images from its own AI models. Meta has said it adds invisible watermarks as well as metadata to images made with its AI features. Adobe and others back “durable” Content Credentials, which pair a C2PA manifest with a watermark or fingerprint so the record can be looked up again. Other generators rely only on metadata, and local tools usually add nothing. Practices change often, so treat any list as a snapshot.
How reliable detection is
Watermark detection is statistical: a detector reports how likely it is that a mark is present, not a yes-or-no certainty. Heavy edits can weaken a mark, and a detector only recognizes its own provider’s watermark. Most importantly, finding no watermark proves nothing — many generators never add one. That is why checking an image for AI means weighing several clues together.
What it means for your privacy
As publicly described, a pixel watermark marks content as generated by a provider’s model rather than storing your prompt as readable text. The details that most often expose you — prompts, account IDs, locations, device data — live in metadata, and that is the layer a washer can clean. In short: washing protects your personal information, while the watermark keeps saying where the image came from.
About SynthID and watermarks
If an image has no watermark, is it a real photo?
No. Many generators add no watermark at all, so the absence of one tells you nothing on its own.
Can a watermark reveal my prompt?
Not as publicly described. Prompts and account details are usually stored in metadata, which washing removes.
Does cropping or a screenshot remove SynthID?
Generally not. SynthID is designed to survive common edits such as cropping, resizing, screenshots and compression.
Does AI Label Remover remove SynthID?
No. It removes metadata only and never edits pixels, so any invisible watermark remains.
How do I know if an image has SynthID?
Only Google’s verification tools can tell you. A metadata reader cannot see inside the pixels.
Is every AI image watermarked?
No. Some providers watermark output, many do not, and local tools usually add none.