AI-generated images can contain several different kinds of identifying or descriptive information. Some is ordinary file metadata that a conventional cleaner can target. Some is provenance data. Some may be encoded into the visible pixels or detectable from the image content itself.
PrivMeta removes documented metadata structures from supported image formats, but it is not a universal AI-watermark or provenance remover.
What Metadata Can an AI-Generated Image Contain?
Depending on the generator and export format, a file may contain:
- EXIF or XMP fields naming software or workflows
- PNG text chunks containing prompts, parameters, model names, seeds, or workflow data
- Comments and application-specific metadata
- Creator or copyright fields
- Content provenance records, including some Content Credentials implementations
- Visible watermarks, logos, signatures, or disclosure labels
- Signals embedded in pixel values rather than stored as conventional metadata
These mechanisms are not interchangeable. Removing an EXIF block does not imply that a visible watermark, provenance credential, or pixel-level signal has been removed.
What PrivMeta Removes from Supported Images
PrivMeta uses different cleaning paths by format.
JPEG and JPG: recognised EXIF, XMP, IPTC, Photoshop resource, and comment segments are removed without recompressing the encoded image. Other JPEG application segments are preserved unless they are explicitly targeted.
Static PNG and WebP: the visible raster is decoded and re-encoded by the browser, producing a new image without the original ancillary metadata containers. Dimensions and visible content are preserved, but byte-for-byte pixels, colour profiles, and exact encoding are not claimed to be identical.
GIF: comments, plain-text data, XMP, and non-animation application extensions are removed while frames and animation timing are retained.
Animated PNG and WebP are rejected because the browser cleaning path would flatten the animation.
What PrivMeta Does Not Promise to Remove
PrivMeta should not be described as removing:
- Visible watermarks or disclosure labels
- Logos, signatures, or text rendered into the picture
- Pixel-level or model-specific detection signals
- Every Content Credentials or C2PA representation
- Arbitrary application-specific structures it does not recognise
- Evidence that might be inferred from the visual content itself
For JPEG specifically, unrecognised application segments are preserved to avoid damaging valid image data. A provenance structure stored in one of those segments may therefore remain.
How to Clean an AI-Generated Image
- Open PrivMeta.
- Add the supported JPEG, PNG, WebP, or GIF image.
- Select Remove metadata.
- Download the cleaned copy.
- Inspect the result with a current metadata tool if provenance or attribution is important.
The cleaning happens in your browser. File bytes are not uploaded to a processing server.
How to Verify the Result
Do not rely only on a filename change or a metadata viewer that reports a short summary.
For a higher-confidence check:
- Run ExifTool on the original and cleaned files.
- Compare reported EXIF, XMP, comments, text chunks, and application-specific groups.
- Compare dimensions and open the cleaned image in the browsers or applications you need.
- If Content Credentials matter, inspect them with a tool designed for that provenance system as well.
A conventional metadata report and a provenance check answer different questions.
AI Metadata Removal Versus Cleaning Before an AI Upload
This page concerns metadata already present in an AI-generated image. If you instead want to protect an ordinary photo, document, audio file, or video before sending it to an AI assistant, read how to remove metadata before uploading to ChatGPT or Gemini.
Use Precise Claims
PrivMeta can reduce conventional metadata exposure in supported files. It cannot make an image anonymous, erase visible evidence, or guarantee that every AI provenance mechanism is absent.
Remove documented metadata from a supported image with PrivMeta, then inspect the result according to the level of assurance your use case requires.