A public body sends a redacted report to a journalist. The names in the text have been properly removed. The document properties name the author, and the file path names the folder the document lived in: the case officer's surname and the matter name.
The redaction did not fail. Only the page was checked.
What is in the file without being on the page
A PDF carries a set of fields you only see if you go looking:
- Author and organisation. Usually picked up automatically from the settings of whatever created the document.
- Title, subject and keywords. Frequently still the working title from an earlier draft.
- Creation and modification dates. Timestamps included.
- Producing software. Which application, which version.
- The file path. Some export methods include the full path of the source file.
- XMP fields. A broader set of fields that applications share with each other, able to hold a great deal, including editing history.
The file path is the underrated one. Something like C:/Users/j.dewilde/Cases/Merger-Northside/draft-v4.docx contains a username, a matter name and a version, none of which appeared on the page.
What comes across from the source
Beyond the standard fields, it depends on where the document came from.
From Word. Margin comments and tracked changes can survive the export, depending on settings. Export a draft with three colleagues' comments in it without accepting or removing them first, and you share those comments. See also what "save as PDF" does not remove.
From a system. Exports from a case or matter management system sometimes carry internal references: case numbers, user IDs, the template name.
From a scanner. Device name, sometimes its serial number and the settings used.
Why redaction does not reach this
This is the part that surprises people, although it is entirely logical once you spell it out.
Redaction operates on page content: this passage must go, so that text is removed. Document properties sit wholly outside that. They are not part of the page, so a tool that edits the page does not touch them.
Which means you can redact a document perfectly, following every rule in "Why a black box over text is not redaction", and still send out the author's name with it.
Removing metadata is a separate action, with a separate button, in virtually every tool.
Why AI tools read this too
When parsing an uploaded file, services generally take more than the visible text. Exactly what gets included varies by provider and is rarely documented.
The safe assumption is therefore: what is in the file goes to the service. Not because anything malfunctions, but because you sent the file rather than the page.
That is also the practical argument for checking metadata before an upload, and not just before you send something outside the organisation. An upload to a chatbot may feel internal, but it is an external transmission.
What to do before sharing
- Open document properties. In any PDF reader, under file or document properties. Look at the author, the title and whether a path is present.
- Run a hidden-information inspection. Most PDF tools have a function that finds metadata, embedded files, comments and hidden layers.
- Remove what has no business being there. Author and organisation rarely serve a purpose in a shared document.
- Check again after editing. Some tools write their own name and a fresh modification date back the moment you save.
Four steps, a minute in total. That is the ratio at stake here: a minute of checking against a piece of data you never meant to share and cannot get back afterwards.
When you anonymise a PDF with BeeSensible, attachments, XMP metadata, JavaScript and application data are always stripped, and afterwards the service checks whether a removed value still occurs in the file.
