Watermarks and Content Credentials provide different kinds of information about an image. A visible label communicates something directly to a viewer. An invisible watermark can carry a signal detectable by a compatible system. Content Credentials can record signed information about an asset's origin and editing history.
None of these should be read as a universal stamp saying that a depicted event is true. Their usefulness depends on what was recorded, which system is checking it and the question you are asking.
Visible labels communicate without special software
A caption such as “AI-generated illustration” tells the viewer how the publisher describes the image. A visible mark can serve a similar purpose inside the picture itself. These labels are easy to understand, but their presence alone does not verify the publisher's account of how the file was made.
For your own publication, choose language that reflects the material changes. A captured photograph with a generated background needs a different explanation from a wholly synthetic scene. Our guide to AI-generated images versus edited photos gives examples of useful distinctions.
Invisible watermarks use a different channel
An invisible watermark is designed to be detected without appearing as ordinary text or a logo. Google DeepMind's description of SynthID explains an image-watermarking approach intended to help identify outputs from a supported generation system.
The scope matters. A check for one watermark is not automatically a detector for every kind of generated image. A negative result should be interpreted using that system's documented limitations, not converted into a universal claim that the image came from a camera.
Also distinguish a detector that searches for a deliberately embedded signal from one that estimates origin using visual or statistical patterns. They answer related but different questions.
Content Credentials describe recorded provenance
The C2PA FAQ describes Content Credentials as cryptographically signed records that can contain information about creation and subsequent changes. Validation can help assess the relationship between the record and the asset and whether the record has been tampered with.
The useful questions are concrete: what does the record say, who signed it, and what does the viewer report about validation? A decorative badge or screenshot of a validation result is not equivalent to examining the relevant record attached to the actual file.
Provenance is also different from permission. A record describing an edit does not, by itself, establish that every person or rights holder agreed to that use.
A missing record does not settle the question
Not every camera or editing workflow creates Content Credentials, and metadata may not survive every transfer. Therefore, a missing record cannot establish that an image is fake or genuine. The C2PA explainer also distinguishes verified provenance from a judgment that the depicted content is factually true.
Consider a hypothetical photograph with a valid creation record but an incorrect caption naming another city. The record can help establish the file's history while leaving the caption wrong. Evaluate the surrounding statement separately.
Use the evidence without overstating it
When publishing your own illustrations, keep the original exports and describe significant generative changes clearly. Preserve available provenance information where the publishing workflow supports it. Do not claim that your files contain a particular credential unless you have checked the actual exports.
When assessing someone else's image, combine its source, context and available technical evidence. If a watermark or credential answers only part of the question, say which part. The guide to reliable AI-image detection explains how to communicate remaining uncertainty without turning a tool result into an unsupported accusation.
Related reading
How AI Image Generation Works: A Beginner’s Guide
AI-Generated Images vs Edited Photos: What Is the Difference?
