Image Metadata, Content Credentials, and Watermarks Explained
Understand EXIF metadata, Content Credentials, visible overlays, and imperceptible watermarks before editing or making claims about an image’s provenance.
When an image moves between a generator, editor, messaging app, and publishing platform, different provenance signals can survive or disappear independently. Understanding those layers prevents claims such as “the metadata is gone, so the image cannot be identified” or “the logo is gone, so SynthID is gone.” Neither conclusion follows.
Visible overlays live in the rendered image
A visible logo or label changes pixels a viewer can see. Cropping, retouching, or regeneration can change those pixels. That may improve composition, but it does not automatically change metadata, signed provenance, or an imperceptible watermark elsewhere in the pixel data.
EXIF and application metadata live in file fields
Traditional metadata can include camera settings, capture date, GPS coordinates, orientation, color profile, creator fields, and the software that last wrote the file. Some export tools retain this information, some selectively remove it, and screenshots often create an entirely new metadata record.
Content Credentials describe provenance and edits
Content Credentials are designed to carry verifiable information about origin and editing history. A compatible viewer may show who signed a credential, what tools were involved, and what edits were declared. Depending on the implementation, provenance information may involve data embedded in the asset and records that are recoverable beyond a single local metadata field.
Imperceptible watermarks are encoded into content
Google’s verification guidance distinguishes SynthID from metadata: SynthID watermarks are embedded directly into content and are not visible to users, while metadata adds context to the original file. Google also notes that a SynthID signal can usually remain after resizing, recoloring, compression, or other alterations.
Detection is not a perfect binary truth machine. A verifier may detect a signal, find no signal, or return an uncertain result. “Not detected” for one system should not be expanded into “definitely human-made,” especially when many AI systems do not use the same watermark.
What a normal image export can change
| Action | Likely effect | What it does not prove |
|---|---|---|
| Crop | Removes edge pixels and changes dimensions | That embedded signals are absent |
| JPEG re-export | Recompresses pixels and may rewrite metadata | That provenance is erased |
| Screenshot | Creates a new file from displayed pixels | That no detector can identify the content |
| Visible mark repair | Reconstructs a localized pixel region | That the image is no longer AI-generated |
| Metadata stripping | Removes selected file fields | That pixel watermarks or external records are gone |
A better provenance workflow
- Keep the original generated or captured file.
- Inspect metadata and Content Credentials with tools meant for those layers.
- Use a compatible detector when you need to ask about a specific invisible watermark.
- Use pixel editing only for the visible image problem you are trying to solve.
- Carry required attribution and AI disclosure into the final publishing workflow.
Last reviewed July 15, 2026. This guide is general product and publishing information, not legal advice.