Provenance

Image provenance verification is a chain, and most chains are missing links.

Provenance is not a yes-or-no property of a file. It is a set of claims, each with its own strength, and the useful work is knowing which claim you actually have before you repeat it.

Four kinds of evidence, in descending strength

Signed provenance is the strongest, and the rarest: a manifest in the file, signed by a known party, with a hash that matches the asset.

Generator and editor tags come next. They identify a tool, not a person, and they are trivially stripped or forged in a re-encode.

Supplier records are the practical backbone: the original file, the export path, the date, an email from the person who made it. They are not cryptographic and they are what most disputes are actually settled with.

Reverse image search comes last. It finds where an image appeared, which is a statement about publication rather than origin.

What a file-level check can say

A browser check reads the bytes and reports which provenance systems and generator names appear in them. That is a genuine finding about the artefact in front of you.

It is not a finding about the world. A missing manifest may mean an unsigned original, a stripped copy or a deliberate removal, and the bytes alone cannot distinguish those.

So the wording matters. Reported as no manifest present is defensible. Asserted as unedited is not.

The stripping problem

Every re-encode is a chance to lose metadata, and the normal publishing path has three or four. Export from an editor, resize in a CMS, upload to a social platform, screenshot into a slide deck.

The practical rule is to capture provenance at the point of creation and keep it beside the asset rather than inside it. A hash plus a manifest copy in the delivery folder survives everything the pixels do not.

For anything used as evidence, keep the original file untouched as well, and publish the derivative. The original is the only artefact that can still answer a question a year later.

When the answer is unclear, say so

The commonest failure in provenance work is a confident statement built on a weak signal: this looks AI-generated, this has no metadata, this appeared nowhere else.

Each of those is a reason to ask a question, and none supports a public claim. Write the finding at the strength the evidence allows and let a named reviewer decide whether to publish it.

That discipline protects everyone. A wrong accusation about a photographer's work is costly, and so is presenting a generated image as documentary.

Provenance as a process, not a check

The organisations that handle this well do four things: they sign at creation where the tool allows it, they keep the original, they record a hash at export, and they store the record with the delivery.

Each step is cheap. Together they convert an unanswerable question into a filing exercise, which is what an auditor actually wants.

Where a supplier cannot provide any of the four, that absence is itself the finding, and it should be recorded as such rather than smoothed over in a deck.

Where a human takes over

Judgement about whether an image implies documentary status, whether a claim in the caption is supported, or whether usage rights cover the placement belongs to a person with the context.

For client work and regulated material, that person should sign a defined scope: reviewed provenance and usage rights for the home page hero, for example.

The check narrows the question. The signature answers it.

What this covers

  • Signed manifest

    Present and signed by a party you can identify, with a matching hash.

  • Tool tags

    Generator and editor names recorded, understood as tool-level evidence.

  • Supplier records

    Original file, export path and a dated statement from the maker.

  • Reverse search

    Prior publication checked, treated as publication evidence rather than origin.

  • Recorded strength

    Findings written at the strength the evidence supports, never stronger.

Questions this page usually raises

Can I prove an image is not AI-generated?
Not from the file alone. You can establish that a manifest is absent, which is common after any resize, or that records exist. Proof of absence is not available.
Is reverse image search reliable?
It finds appearances. A generated image with no prior publication returns nothing, and so does an original photograph, so an empty result distinguishes nothing.
What should we ask a supplier for?
The original file, the export path, the date and the tool used. For generated assets, the signed output where the tool supports it.
How do we keep provenance through a publishing pipeline?
Sign at creation, keep the original, hash at export, and store the record with the delivery. Serialising the pipeline is the other option, though rarely practical.
Does the check require the image to be uploaded?
No. Everything runs locally in the browser, which is what makes it usable on unreleased and client material.

Image provenance verification

Run the checks, then decide who has to sign it.

The automated pass is free, needs no account and uploads nothing. Escalate to a named reviewer only where being wrong is expensive.