What provenance can and cannot tell you

Signata answers two questions: where did this come from? and has it changed since it was signed? It does not answer is this true? or is this fake?, and we have tried hard to make sure nothing in the product can be read as if it does.

Absence of provenance is not proof of fakery

Most legitimate content on the internet today carries no Content Credential at all. Cameras without the feature, screenshots, scans, older files, anything re-saved by a tool that doesn’t write credentials. All of it is unsigned and perfectly genuine. So no_provenance means exactly one thing: we cannot tell you where this came from. It is not a verdict of “fake,” “manipulated,” or “AI.”

A valid signature proves origin and integrity, not truth

When a verification comes back verified, two things are true: the credential was signed by the holder of a key you trust, and the asset has not changed since it was signed. That is real and valuable. But it says nothing about whether the content is honest, accurate, or undoctored-before-signing.

A perfectly genuine, trusted signature can sit on a staged photograph, a misleading crop made before signing, or an image whose caption lies. The signature vouches for the pipeline, not for reality.

Signers can assert false things within a valid signature

A signature binds a claim to a key; it does not police the claim’s contents. A signer can record an inaccurate edit history, an untrue author, or an asset they didn’t actually create, and the result will still verify, because the cryptography only attests who said it, not whether it is so.

This is exactly why trust is a deliberate, separate decision. Trusting a signer means you have reason to believe that party tells the truth in its credentials. An untrusted-but-valid credential is reported honestly as such:

Metadata is easily stripped in transit

Content Credentials live in the file. Many platforms re-encode uploads or strip metadata, which removes the embedded manifest, so a credentialed asset can arrive looking like it has no_provenance. We mitigate this with the transparency log and soft binding: provenance can often be recovered by content hash.

But recovery is best-effort, not guaranteed. It only works for credentials Signata issued, and only while the bytes still hash to the same value. Re-encoding changes the hash, and the link is lost. Stripping is a real limitation, partially addressed, not solved.

Trust lists are your policy choice

Whether a signer is “trusted” is not a fact about a signature. It is a policy you set. Signata ships known issuers (Adobe, OpenAI, supported cameras) as a convenience, but adding any of them to your trust list is an explicit choice, never an automatic grant. Two organizations can verify the same asset and reasonably reach different verdicts because they made different trust decisions. That is by design.

Signata is provenance, not deepfake detection

Signata does not analyze pixels to guess whether an image is AI-generated or manipulated. It reads cryptographic credentials and checks bindings. AI involvement is surfaced only when it is disclosed in the manifest, never inferred. If you need to detect undisclosed manipulation in unsigned media, that is a different tool and a fundamentally less certain one; we would rather be narrow and honest than broad and wrong.