An image label can tell you something about a file’s history. It cannot tell you whether the picture is true.
That distinction matters when a photo is being used to sell something, support a claim or provoke a reaction. A label can be useful evidence. It is not a truth machine.
Content Credentials, the C2PA standard’s name for signed provenance data, can record claims about an asset’s origin, edits and use of AI. The signature and hash let a viewer check whether that record still matches the file. C2PA is explicit about the limit: validation checks that provenance information is well-formed, untampered with and associated with the asset. It does not decide whether the account written in the record, the scene in the image or a caption beside it is true.
Use the label for the question it can answer
Start with the actual image or video file when possible, not a screenshot or a copy downloaded from a social feed. Services can remove metadata while resizing or re-encoding media. The C2PA FAQ says embedded manifests can be separated from their files, although some systems use watermarking or fingerprinting to help recover a credential later. So a missing credential is not proof that no AI was used. It may mean the file never had one, the platform discarded it or the viewer cannot find it.
Here is a practical decision tree:
- A viewer reports a valid Content Credential: read the specific creation and edit actions. Treat it as evidence about the documented file history, then separately check the claim the image is being used to support.
- A viewer reports a problem with the credential: stop treating that record as reliable for this version of the file. Find the original from the publisher or creator, if one is available.
- No credential is found: record only that result. Do not turn it into “human-made” or “fake.” Check the source, date, location and independent reporting instead.
A product-specific watermark is another, narrower signal. Google’s SynthID documentation describes an invisible watermark designed to identify content made or altered by Google AI and to survive common changes such as cropping, filters and compression. A detected watermark supports that limited finding. An absent detection does not clear content made by another tool, or content that was edited after a detector’s coverage ends.
A worked example
Hypothetical example: an image of a flooded street is shared with the claim that it shows a particular town this morning. A valid credential says the file was created with generative AI. That is enough to reject the location claim as evidence from a camera, but it does not tell you why the image was posted or whether any real flooding happened. Now change the result: no credential is found. The correct next step is still not to trust the caption. Look for a dated report from a local authority or news organization, and compare visible landmarks with reliable material. The missing label settled nothing.
If you need to inspect a file, use a viewer that explains whether it found a credential and whether validation passed. Adobe’s Inspect documentation describes a file-upload inspection tool and says uploads are not stored. That statement applies to that tool, not every checker, so read the privacy terms before uploading sensitive media.
The finish line is modest but useful: you can state exactly what the label establishes, what it does not establish and what independent evidence would answer the remaining claim. A good provenance record narrows the work. It does not let anyone skip it.