SynthID and Content Credentials are two different marks

When an image from an AI tool gets a label or a "CR" icon, people tend to treat it as one thing: "the AI mark". In practice there are at least two unrelated mechanisms, and they behave differently when you edit or clean a file. Mixing them up leads to a lot of wrong expectations, so here is how I keep them apart.

Mark 1: Content Credentials (C2PA) live beside the pixels

Content Credentials are the user-facing name for C2PA data: a signed manifest attached to a file. It records things like which app or device made the media, who issued the credential, when, and what the signer says happened (captured by a camera, edited, or generated by AI).

Key properties:

It is metadata. It sits in the file container, next to the image data rather than inside it. In a JPEG it typically lives in APP11 segments using JUMBF; in a PNG, in a caBX chunk.
It is a statement by whoever signed it. It is not a measurement of the pixels, and it is not only for AI: cameras and news organizations use it too.
It can be removed by rewriting the file without that segment, or lost by re-saving through software that doesn't carry it over.
A signer can also publish a copy to a cloud service. Removing the attached manifest doesn't touch a copy stored elsewhere.

LinkedIn, for example, shows its CR icon when an uploaded image or video has C2PA information attached. Open the icon and you see fields such as the app or device, the creator, the issuer, and the date.

Mark 2: SynthID is a watermark in the content

Google describes SynthID as a watermark embedded into AI-generated content. Google's own help text for Gemini draws the line explicitly: SynthID is the invisible watermark, and Content Credentials are the C2PA metadata about origin. A single file from a Google tool may carry both.

Key properties:

It isn't stored in a tag you can list. There is no field called "SynthID" to look up in an EXIF dump.
It was designed to be invisible to people, and detection needs the right detector, not a metadata viewer.
Because it isn't metadata, tools that work on metadata have nothing to delete there.

What a metadata cleaner changes, and what it doesn't

If a tool strips EXIF, GPS, IPTC, XMP and C2PA, then after a clean:

EXIF (camera, GPS, timestamps): Yes, if the mode includes it
IPTC / XMP fields, e.g. a "created by AI" source type: Yes, if the mode includes them
C2PA manifest attached to the file: Yes
A copy of the credential published to a cloud service: No, it isn't in the file
A watermark such as SynthID: No, it isn't metadata
A sentence in your caption saying the image is AI-generated: No, it's text you wrote

That last pair is the part people miss. If the goal was "make the file carry no AI provenance data", a metadata clean may do that. If the goal was "make it impossible for anyone to tell", no metadata tool can promise that, and I'd be suspicious of any tool that says it does.

A note on disclosure

Some platforms and some jurisdictions expect AI-generated media to be disclosed. Removing a manifest from the file you upload is not the same as disclosing, and it isn't a substitute. LinkedIn's own help says it's not yet possible to identify and label all AI-generated and modified content, so a missing icon proves nothing either way. Use cleaning for legitimate reasons, like not leaking GPS coordinates or a drafts folder's tags, and follow the disclosure rules wherever you publish.

How to check what a file has

Run exiftool -a -G1 -s yourfile.jpg and read the EXIF, IPTC and XMP groups.
Look for a C2PA / JUMBF container with a C2PA-aware viewer or a small script that scans the file's segments.
Remember the absence of a C2PA line only tells you about that file. It says nothing about pixel watermarks.

I maintain a browser-based cleaner that only handles the metadata side and says so on the page. If your question is specifically whether a tool like that can take out SynthID, this page gives the short answer and the reasoning: no, because the cleaner strips tags and doesn't edit pixels.

Summary

Content Credentials: signed metadata, removable, a statement by the signer.
SynthID: a watermark in the content, not listed as a tag, not removed by metadata cleaning.
A file can have both; a cleaning tool addresses only the first.
Neither one replaces a disclosure you owe a platform or an audience.

Not affiliated with Google or LinkedIn. Descriptions are based on their public help pages as of late September 2026.

More notes in this series: What metadata is inside a Midjourney download? Check before you publish · Before you deliver photos to a client: check for GPS, camera serials and AI edit history · What is the "CR" icon on LinkedIn images, and why is it on my photo? · How to check whether a JPEG or PNG carries C2PA data (with a short script)