Transmission // 08
Nobody signed the picture
3 min
Content Credentials became the obvious way to satisfy the EU's labelling rules. It still cannot tell you anything about an unsigned file — and a signed one is not proof either.
C2PA — Content Credentials — does something narrow and does it well. It binds cryptographically signed provenance metadata to a media file: who made it, when, with what tool, whether a model was involved, and every meaningful edit since capture. Its AI assertion type maps cleanly onto the EU AI Act’s Article 50 labelling duty that became enforceable in August, and a conformance programme has been pulling in implementers since 2025 — Google in August of that year, the IPTC in April 2026.
Regulatory pull is the strongest adoption force this standard has ever had. It is still not enough, for two reasons: one structural, one that has already happened.
The asymmetry that breaks it
A valid credential proves something. A missing credential proves nothing.
Absence of provenance is the default state of essentially all media in circulation. It is also what you get from: a screenshot, a re-encode, a platform that strips metadata on upload, a phone from 2019, a scanner, a fax, a crop in a tool that never joined the coalition, or an attacker who simply did not sign.
So the verification question can never be “is this credentialed” — it is “is this credentialed and did I have a prior expectation that it would be”. That prior only exists inside closed pipelines: a newsroom’s own cameras, a studio’s own render farm, an insurer’s own claims app. Inside those, C2PA is genuinely excellent. Outside them, it is a green tick that most true things will never earn.
And a valid credential proves less than it looks like
In August 2025 a Nikon firmware update exposed a flaw that let a user combine unauthentic images with authentic photos and still produce a file carrying a valid signature. Nikon revoked the certificates.
That is the failure mode that should worry anyone treating the green tick as an answer. The signature attests that a specific device asserted a specific history. It does not attest that the history is true — only that the device said so, and devices are firmware, and firmware has bugs. Chain of custody is a statement about process integrity, and process integrity is exactly what an attacker attacks.
What it is actually for
Stop thinking of it as a detector. It is a chain of custody, and chains of custody have always worked the same way: they raise the cost of a specific lie inside a specific system, and say nothing about the world.
That is worth building. It is not worth pretending it is deepfake detection, because the failure mode of that pretence is the dangerous one — a public trained to read “no credential” as “fake” will disbelieve every real photograph taken by anyone without an enterprise imaging workflow.
The version that would work
Signing at capture, in hardware, by default, in every phone — with the signature surviving the ordinary edits people actually perform. That is a decade of platform coordination and a set of privacy questions nobody has answered, since a signature that proves provenance also proves presence.
We will get the labelling requirement first and the infrastructure later. That order is backwards, and it is the order we are going to get.