2026-08-12

Does removing a photo's metadata also strip its AI Content Credentials (C2PA)?

A metadata scrubber and an AI-content authenticity standard are trying to do almost opposite things to the same file, and as of this month that's no longer a theoretical collision. One strips identifying information out of a photo before it leaves your device. The other cryptographically signs a photo so a viewer can later confirm nothing about it has changed since. NearScrub only does the first of those — and knowing exactly where the second one lives inside a JPEG explains why running a photo through any metadata tool, NearScrub included, has a side effect worth knowing about before it matters.

Where a Content Credential actually sits inside a file

C2PA — the Coalition for Content Provenance and Authenticity, the standard behind what Adobe brands "Content Credentials" — doesn't store its manifest as EXIF or XMP, even though it's often lumped in with "photo metadata." Per the C2PA technical specification, a JPEG's Content Credentials are packaged as a JUMBF box (JPEG Universal Metadata Box Format, an ISO/IEC 19566-5 standard) and embedded in the file's APP11 marker segment — a different marker entirely from APP1, where EXIF and XMP live, and APP13, where Photoshop/IPTC data lives. PNG files carry the same JUMBF structure inside a dedicated caBX chunk, kept separate from the tEXt, iTXt, and other text chunks a PNG uses for ordinary metadata. The standard was built that way on purpose, so a signed manifest survives tools that only know how to touch conventional metadata fields.

Why "the manifest survives" still doesn't mean "the credential still verifies"

That separation only protects the manifest's own bytes, not what it certifies. C2PA's own FAQ (c2pa.org) is direct about this: "Any modification—intentional or accidental—will break this cryptographic linkage, signalling tampering." The mechanism is what C2PA calls a hard binding — a cryptographic hash, taken at signing time, covering the asset's bytes outside the manifest box. Change anything in those bytes afterward — crop the image, resave it at a different quality, or run a scrubber that blanks the EXIF and XMP segments — and the hash no longer matches what the manifest recorded. A verifier doesn't then report "no credential found"; it reports the credential as invalid, which reads less like "the metadata was cleaned" and more like "this file was altered after signing." For a photo whose whole point is proving it wasn't manipulated, that's a meaningfully different message to end up with.

Why this stopped being an edge case on August 2, 2026

Article 50 of the EU AI Act's transparency obligations became applicable on August 2, 2026 — ten days before this was written. The European Commission's Code of Practice guidance describes the compliance approach it's steering providers toward as "a revised two-layered marking approach involving secured metadata and watermarking" for AI-generated outputs. C2PA is the leading open standard built for exactly that machine-readable metadata layer, and its adoption curve backs that up: the Content Authenticity Initiative — the C2PA-aligned industry group Adobe founded — passed 5,000 member organizations in 2025, up from 4,000 the year before, according to its own blog. Its 2025 milestones read like a checklist of the capture-to-delivery chain now embedding manifests by default: Leica's SL3-S and Samsung's Galaxy S25 line shipped with native Content Credentials at the point of capture, Sony's PXW-Z300 became the first camcorder to sign C2PA video, and Cloudflare became the first major CDN to preserve Content Credentials instead of stripping them on the way through. A generative-AI output or a camera-captured photo increasingly arrives already carrying a manifest, which means a metadata tool run on it downstream isn't acting on a blank slate anymore.

Two different jobs, one file — and they usually can't both be done

None of this makes metadata scrubbing the wrong move; it makes it a trade-off worth naming before it's discovered by accident. If a photo's job is proving where you were, GPS metadata is exactly the risk a tool like NearScrub exists to remove, same as always. If a photo's job is proving its own authenticity to whoever receives it — a newsroom, a marketplace requiring provenance, a platform complying with Article 50 — the file needs its Content Credential to keep verifying, and running it through any metadata tool first, NearScrub included, breaks that verification instead of quietly cleaning the file around it. There's no version of "scrub the personal metadata but leave the signed authenticity manifest intact" available from a tool that touches the file after signing — the hard binding exists specifically so that option can't exist.

What NearScrub actually touches, and what it leaves alone

NearScrub's JPEG handling drops the APP1 Exif segment, the APP1 XMP segment, APP13 IPTC/Photoshop data, and COM comments, entirely in the browser, nothing uploaded. It doesn't parse APP11 at all, so a JUMBF/C2PA box passes through as untouched bytes if one is present. That's not the same as preserving the credential: because the surrounding EXIF and XMP bytes did change, the manifest's hard binding no longer matches, and a Content Credentials check will flag the result as invalid rather than as absent. If a file needs to keep verifying — because it's headed somewhere that checks — don't run it through NearScrub, or any other metadata tool, first. If a file's job is simply to stop carrying your location and device details before you send it, that's exactly the case NearScrub is built for, C2PA manifest or not.