Image checks and content provenance

Pixels can tell a story. Provenance can tell you where the story came from.

Pixels are not the whole story

An image can be authentic and misleading. It can be edited without being wholly fabricated. It can be real but placed beside the wrong event or date. Image verification therefore needs more than a question about whether pixels were generated by a model.

Start with context. Where did the file first appear? Does the scene match the claimed time and place? Are there older versions, a wider crop, or a related report? Reverse-image search can find reuse, but a match is a lead, not a verdict.

Write down the claim attached to the image before inspecting it. “This photograph was taken here yesterday” makes claims about place and time. “This image shows damage” makes a claim about what the scene depicts. A visual review may address one part and leave the other open. Separating them avoids treating a genuine file as proof of a false caption.

Provenance adds a trail

Content provenance records actions around a file. The C2PA specification describes a standard for signed claims about creation and edits. A compatible credential can help show who asserted what, and when. It can support a chain from capture to publication.

That chain has limits. A credential can be removed by conversion or cropping. A valid signature does not prove that the depicted event happened. It proves something narrower about the record and its signer. Lack of a credential does not prove manipulation. Many ordinary workflows do not preserve one.

Provenance statements are assertions with an issuer and a scope. A camera may assert that a file was captured by a device. An editor may assert that a crop was made. A publisher may bind a caption to a version. These are useful statements when their meaning is clear. They should not be expanded into claims the credential never made. A signed chain is evidence about handling, not a universal authenticity stamp.

Inspect the content and the container

File metadata can offer useful clues. It can also be stripped or changed. Treat it as one input among several. Compare visual details. Look for lighting, reflections, shadows, typography, and repeated patterns. Do not turn an artifact into a conclusion without corroboration.

Compare the image with nearby frames when they exist. Look for landmarks, weather, shadows, and signs that can be checked independently. Be careful with visual “tells.” Compression artifacts and unusual faces can come from ordinary editing, resizing, or a difficult camera condition. Detection tools are indicators. They are not reliable proof in every setting.

The NIST AI Risk Management Framework is a useful reminder to map risks to context. A newsroom, a platform, and a private researcher may need different evidence and escalation rules.

Context should be checked outside the file as well. Compare the scene with maps, public schedules, weather records, or contemporaneous reporting when those sources are appropriate. Each source has its own error and timing. Agreement helps, but it does not remove the need to explain what was actually compared.

What a responsible check says

A responsible image check separates observations from inferences. It says what the file contains, what the available provenance supports, and what remains unknown. It links to the original or earliest reliable source when possible.

It should also preserve the item being reviewed. Record the URL, download time, file hash when practical, and any transformations made for analysis. Do not overwrite the original with a marked-up copy. Keep notes about which tool produced which signal. These habits make a later correction possible.

Prove can help organize this work. Use it to inspect a claim and its supporting material. For important decisions, keep a human in the loop and preserve the file, context, and date of review.