What is content provenance, and what is C2PA?
A way of recording where a piece of media came from and what was done to it, cryptographically attached to the file — an approach to the synthetic media problem that works in the opposite direction to detection.
Why detection is the wrong strategy. Identifying AI-generated content after the fact is an arms race that the detectors are losing: generators improve, detectors produce false positives on genuine material, and any detector becomes a training target. Detection fails quietly and asymmetrically.
The provenance approach instead. Rather than trying to identify fakes, record and verify the history of genuine content. The question shifts from "is this fake?" — which is hard — to "what can this file prove about its own history?" — which is tractable.
What C2PA is. The Coalition for Content Provenance and Authenticity, a standards body including technology companies, camera manufacturers and news organisations, which has produced an open technical specification.
How it works. A manifest is attached to the file recording assertions — the capture device, the time, the software used, and each edit applied. The manifest is cryptographically signed, so tampering is detectable, and each subsequent edit adds to the chain rather than replacing it.
Content Credentials is the user-facing name, displayed as an icon that can be inspected.
What it does and does not prove:
It proves the chain of custody, not that the content is true. A signed photograph of a staged scene is still a staged scene.
Absence proves nothing. Most content will not carry credentials, so a missing manifest cannot imply fakery — and this is the crucial limitation, since a system whose absence means nothing is only useful where it is present.
Stripping is trivial. A screenshot removes it, and most platforms have historically stripped metadata on upload — though several now preserve it.
Watermarking is the complementary approach: embedding an imperceptible signal in the content itself that survives some editing. More robust to stripping, less informative, and subject to its own adversarial pressure.
Adoption is the determining factor, and requires cameras, editing software and platforms all to participate.