What is C2PA?
The Coalition for Content Provenance and Authenticity defines a way to attach a signed record of origin and edit history to a media file. It is the single strongest input to an automatic AI label - and the least understood.
What a manifest actually contains
A C2PA manifest is a signed structure stored inside the file: in a JUMBF box for JPEG and PNG, in a uuid box for MP4-family video. It can record which tool produced or edited the asset, when, under which account, and a list of assertions about what happened - including whether a generative model was involved.
Because it is signed, a platform can check that the record was not tampered with. That is the point of the standard, and it is also why the record is so easy for a platform to act on: no analysis of the picture is needed, only a read of the file.
Which platforms read it
- Meta (Instagram, Facebook, Threads) - shows AI Info when it finds AI-related provenance.
- TikTok - applies an automatic AI-generated label, which its own guidance says cannot be removed from a published post.
- Pinterest - marks pins as AI modified, and lets users see fewer of them, so the effect reaches distribution rather than just display.
- LinkedIn - renders a Content Credentials badge on images carrying a manifest.
- YouTube - reads provenance alongside its own disclosure questions for altered or synthetic content.
Why a real photograph gets flagged
The manifest describes an edit, not a proportion. Use Generative Fill to remove a bin from the corner of a photograph you shot yourself and the export carries a credential describing that operation. A platform reading it sees AI involvement and labels the post. Nothing about that is a bug - it is the standard working as designed, applied to a case its wording does not fit well.
Meta renamed its badge from Made with AI to AI Info for exactly this reason. The rename softened the wording; it did not change what triggers it.
What removing it does, and does not do
Removing the manifest removes the metadata-driven trigger. It does not touch a pixel watermark such as SynthID, it does not affect a visual classifier, it does not change a post that is already published, and it does not replace a disclosure you are required to make. Any tool that suggests otherwise is describing a product that does not exist.
Four layers sit between your picture and the label
A tool that treats all of them as metadata is selling false confidence. Here is exactly which two we address, and which two nobody can address from a browser.
Visible pixels
What a person sees. Crops, filters, burned-in logos and subtitles. Changing them proves nothing about the hidden layers.
Pixel watermarks
SynthID and similar systems embed a signal into the image data itself, designed to survive editing. No browser tool removes these, including this one.
C2PA content credentials
A cryptographically signed record of origin and edit history. This is the strongest single trigger for an automatic label.
EXIF, XMP, IPTC and text
Ordinary file metadata: camera fields, GPS, editor names, prompts, seeds, workflow dumps and the IPTC field that marks an image as algorithmic.
Frequently asked questions
Is C2PA the same as a watermark?
No. A C2PA manifest is metadata attached to the file: remove it and it is gone. A watermark such as SynthID is embedded in the pixels and survives ordinary editing. They are frequently confused, and the confusion is why people think a metadata tool failed when it did exactly what it claimed.
Can C2PA be verified after I remove it?
No. The manifest is what carries the signature, so removing it removes the ability to verify anything. That is the trade-off: no provenance claim, and no automatic label derived from one.
Do all AI tools write C2PA?
Most large ones do now - OpenAI, Adobe and Google among them. Others write XMP or IPTC fields instead, which platforms read as well. Which fields your export carries depends on the tool and its version, which is what the checker is for.
Is removing content credentials legal?
Working on files you own is ordinary file handling, and metadata removal is a standard part of publishing workflows. What is not optional is following each platform's rules and any disclosure obligation that applies to genuinely generated content. Nothing here replaces that.
When should I keep C2PA?
Photojournalism, evidentiary work, client deliverables where edit history is part of what you deliver, and any workflow with a compliance requirement. Provenance has real value; the point is to decide rather than to have it decided for you by whatever tool touched the file last.
See whether your file carries a manifest
The checker reads the container and reports what is there. It modifies nothing.