I have this thought, but not sure if it can be considered a decent solution to achieve machine blindness?
One of the reasons why PRISM/computers or search works is because they can parse the text we type off our keyboards. Those character codes is what helps 'em look for key words, and enables the opportunity to snoop.
So instead if we doodle our text [1] in free-form and in language/lingo of our choosing (augment steganography with raw data/images) then we are perhaps closer to making it difficult for the snooping algorithms to parse and detect the keywords these people bank on.
I have significant reasons to believe that OCR across all languages for 400BB mails a day is next to impossible with available technology today. Given the diversity of handwriting/languages worldwide and the the fact that people can invent characters to communicate with each other, this could become an impeccable method.
[1] Example: https://bubbleideas.com/letters/the-tiger-by-sir-william-bla...
(Observe the second half of the page that is handwritten. Since private mails are meant to be read only by intended recipient, we don't need to fret much about search. Basic tagging would be enough.)