Who Says Who's Trusted?

Content provenance has a quiet problem, and the specifications are honest about it.

The cryptography works. A signed asset can be checked. A verifier confirms that the bytes are unchanged, that a signature is valid, and that a particular key produced it. All of that is settled engineering, and the standards bodies have done careful work.

Then comes the question the mathematics cannot answer. The signature belongs to somebody. Should you believe that somebody?

The standard says the question is open

The Creator Assertions Working Group defines how an identity assertion binds a named actor to a signed asset. The identity claims aggregation credential is the mechanism, and the specification is precise about the cryptography: the credential is secured with a COSE signature, the issuer is identified by a decentralized identifier, and the issuer’s public key must be discoverable through a publicly documented DID method.

A word on that first requirement, because the choice carries more meaning than the acronym suggests. COSE — CBOR Object Signing and Encryption, defined in RFC 9052 — is a compact binary format for wrapping a piece of data together with a signature over it. Think of an envelope. The credential goes inside as the payload, the signature is computed across the envelope’s contents, and a verifier opens the envelope and checks the signature against exactly the bytes it finds there.

The important part is the word exactly. The specification states that the unsecured credential is the payload, unmodified. Nothing is reformatted, reordered, or looked up on the network before the signature is checked. Some credential formats work the other way: they transform a document into a normalized form first, which can involve fetching definitions from remote servers, and the signature covers the transformed result rather than the bytes in front of you. Those formats exist for good reasons, but they make what a signature covers depend on something outside the document. CAWG chose the simpler and stricter path. What you verify is what you were given.

Then the specification says this, in the section on validating such a credential:

As of this writing, the Creator Assertions Working Group has not provided guidance about which identity claims aggregators to consider trusted.

Validators are instructed to maintain their own lists. One list of issuers considered trusted. One list of trust anchors considered trusted. The specification defines exactly how to check a signature and explicitly declines to define whose signatures deserve checking.

Look at how implementations handle the gap and the picture gets sharper. The reference Rust implementation compares the issuer’s identifier against a configured allow-list using exact string matching. That is a reasonable engineering decision. It is also a complete description of the state of the art: a signer is trusted because someone put a string in a configuration file.

Why the gap matters more than it sounds

Provenance is often described as proof of origin. In practice a verifier produces two very different kinds of statement, and the difference is easy to miss.

The first is a fact about bytes. This file has not changed since signing. The signature is valid. Nobody has to be trusted for that statement to hold, and the statement is genuinely valuable.

The second is a claim about the world. This was published by a particular news organization. That claim is not established by cryptography at all. Cryptography establishes that whoever held a key made the claim. Everything else — that the key belongs to that organization, that the organization is what it says it is, that the organization has any standing worth caring about — comes from somewhere outside the signature.

Most provenance products answer the second question by answering it themselves. The vendor checks the customer, issues the credential, and displays the result in the vendor’s own verification tool. Nothing about that is dishonest, and the checks are often real. But the answer to “who says this publisher is real?” is “the company selling the provenance service,” and a reader has no way to evaluate that judgement, because the judgement was never published.

For a reader, then, “verified” means the signer appeared on a list. Which list, maintained by whom, on what evidence, is usually not visible.

Accreditation is the part that already exists

Journalism has spent years building the missing piece, in a form that has nothing to do with cryptography.

The Journalism Trust Initiative assesses news organizations against a published standard covering ownership transparency, editorial independence, corrections policy, and sourcing practice. Reporters Without Borders and similar bodies do related work. The assessment is performed by an accreditation body, against public criteria, with a process anyone can read.

An accreditation is a different kind of statement from a vendor’s identity check, in one specific way: the authority behind the statement can be evaluated independently. You can read the standard. You can decide what the certification is worth. You can disagree with the accreditation body’s judgement and say why. A vendor’s internal verification offers none of that, not because vendors are careless, but because an internal process has no public criteria to argue with.

So the useful move is to carry the accreditation as a credential — issued by the certification body, bound to an identifier the publisher controls, and resolvable by anyone. Then a signed article carries two things at once. The provenance of the work: who wrote it, what changed, where the components came from. And the standing of the publisher: an accredited news organization, with the accreditation traceable to the body that granted it.

Those two together are the signal a person or an application can actually act on. Either one alone falls short. Provenance without standing tells you a file is intact and leaves you to guess who made it. Standing without provenance is a badge on a web page.

What this does and does not solve

Worth being precise, because trust models attract more claims than they deserve.

Publishing standing as a resolvable credential does not eliminate trust. Trust cannot be eliminated. Somebody always vouches. What changes is who does the vouching and whether their reasoning is inspectable. Moving the judgement from a provenance vendor to an accreditation body relocates trust to an organization whose entire purpose is making that judgement in public, against criteria it has published.

It also does not settle the standard’s open question. Deciding which identity issuers to trust remains, correctly, a decision for whoever operates a verifier. What an accreditation credential provides is evidence to make the decision with, and a shape for the answer that does not require every verifier to trust the same commercial party.

The cryptography in this field is finished. The interesting work left is governance: publishing who vouches for whom, on what basis, in a form other people can check. Until that work is done, “verified” is a word doing considerably less than readers assume.