> something a machine can do with 100% precision

Forget about typoes. Until recent LLMs, machines couldn't detect oblique or identifying references. (And with LLMs, you still have the problem of hallucinations. To say nothing of where you're running the model.)

> if you have to waste the attention of such people on doing ink redactions instead of dedicating all of their time to focused reading

You've never read a text with a highlighter or pen?

Out of curiosity, have you worked with sensitive information that needed to be shared across security barriers?