The most consequential mistake in this whole category is drawing a black rectangle over a name and believing it has been removed. It has not. The text is still in the file, sitting underneath a shape, and anyone who selects the area and pastes it somewhere gets the original wording back. This has produced real, published, professionally embarrassing failures, repeatedly.
So this tool does the other thing. The characters are deleted from the file. Afterwards there is nothing behind the marks, because there is no “behind” — the words are not in the document any more, not in the search index, and not recoverable from it.
Two design choices follow from that, and both are deliberate. The first is that marked text is outlined, not blacked out, while you are working. A black box is the visual language of the covering tool, and the two must never be confusable at a glance. The second is the sweep: after you mark something, it looks for every other place that looks like it. A name in a witness statement appears in the first line, the signature block, and the address at the bottom, and removing one of three achieves nothing. The sweep finds the others and asks you to confirm them, rather than assuming.
Because it cannot be undone, it confirms before it acts, and it lists exactly what is about to go. Before you download you can always start again from the file you opened; after that, the copy you made is final.
There are two things this cannot do, and it says so at the time. Scanned pages have no text to remove — only an image — so covering is the only option there. And hidden details like the author’s name, the creation date and the original file path live outside the page content; cleaning those out is a separate tool, and worth running before a document leaves your organisation.