What can be removed from a PDF?

Free AI Wiper reads supported document-information fields, such as author, title and creation details, plus XMP metadata attached to the catalog, pages and other indirect PDF objects. Supported embedded JPEG image streams are also checked with the image cleaner. It reports a recognized AI marker separately from ordinary document metadata. You choose whether to remove detectable AI metadata, personal metadata, or both.

For example, a PDF may have an author field even when that name is nowhere on the visible page. Removing the field changes the file's metadata, not the printed text. The image cleaner uses a different method because images store metadata in different containers.

Check, clean, check again

  1. Open the cleaner and choose a PDF of up to 25 MiB.
  2. Review the findings before selecting what to remove. A field without a recognized AI marker is ordinary document metadata, not proof of AI generation.
  3. Download only after the output check shows the selected supported fields are gone.

An already-clean file is not rewritten just to produce a download. If you need to keep an audit trail, retain the original alongside the cleaned copy.

Embedded images and object metadata

A PDF can retain creation metadata inside a picture even after its document properties are cleared. Supported JPEG streams using a single DCTDecode filter are cleaned without re-encoding their picture data. Object XMP uses uncompressed XML or a single FlateDecode filter. The output check compares protected PDF objects and image bytes; a metadata reference shared with document content is refused.

Processing is bounded to 25,000 indirect objects and 128 XMP attachments. Each XMP stream is limited to 1 MiB encoded and decoded, with 4 MiB decoded overall. Supported embedded JPEGs are limited to 128 images, 25 MiB each and 50 MiB together. The PDF input limit remains 25 MiB.

Processing limits

A small PDF can expand substantially when opened. Before normal loading, the cleaner checks supported compressed cross-reference and object streams, limiting decoded data to 8 MiB per stream and 16 MiB overall. Excessive structural entry counts and unsupported structural encodings are refused with an error. These checks reduce exposure to oversized compressed structures; they are not a guarantee that every PDF will process smoothly on every device.

Ordinary supported compressed PDFs can be rewritten while their decoded document objects and content remain verified.

Where this stops

Encrypted PDFs and PDFs with digital signatures are rejected. Cleaning a signed PDF would invalidate its signature. The tool does not remove a visible watermark stamped onto a page or rewrite document text. Inline images, non-JPEG image streams, filter chains, custom decoding parameters, attachments and metadata in uninspected locations remain outside this cleanup. Unsupported XMP encodings are refused. A clean report means the supported fields selected for removal were not found on reinspection.

For the difference between file metadata and a label shown by a platform, see why an AI label may remain.