AI Metadata vs Invisible Watermarks: What Is the Difference?
When someone says an image carries an AI-related signal, they may be referring to very different technologies. Some signals live in metadata containers, while others can be embedded in the image content or pixels themselves.
AI metadata: readable file-level data
Information about a generation or editing tool can appear in XMP, C2PA, text fields or other metadata containers. This information is separate from the visible picture and can be read from the file structure.
Rebuilding a file or removing supported containers can remove this kind of data, which is why a cleaner can rescan the output and compare it with the source.
Invisible signals in the pixels
Other technologies attempt to encode a signal in the image data itself instead of relying on removable metadata. That layer does not necessarily disappear when EXIF or XMP is deleted.
For that reason, a metadata tool should not claim that it removed every invisible watermark or every method that could be used to identify an image source.
What does this mean before publishing?
For privacy, metadata inspection is useful for removing location, device and unnecessary workflow fields. For provenance, it is important to distinguish file metadata from pixel-level signals.
A platform can use more than one signal when applying its own policies, so file cleanup is a specific technical step rather than a guarantee about an external classification decision.
Frequently asked questions
Does MediaClean remove SynthID?
We do not make that claim. SynthID represents a different layer from ordinary metadata, while the current MediaClean workflow focuses on readable and removable file data.
Does re-encoding guarantee every signal is removed?
No. Re-encoding changes the file structure and can affect some data, but it does not guarantee removal of every watermarking or classification system.