The gold standard for redaction is to replace the redacted text with nonsense of similar length. Otherwise, you retain precise metrics for the redacted text. With guesses or context you can fill those in with high probability. In an extreme case, the redaction might be a name of let's say a member of Congress, so the candidates can be narrowed down to a tiny number. There's considerable work on this, notably the "Declassification Engine" [0]. I believe it would be possible to apply even better word models (such as GPT-2) to improve the results even more.

I'm interested in whether this document was redacted in such a military-secure way, or whether black bars were simply placed over the text. I've reached out to a couple of news organizations offering my consulting help, but didn't get a nibble.

[0]: https://www.newyorker.com/tech/annals-of-technology/the-decl...