(Actually, that would be an interesting experiment - remove or replace fraction x of the letters in a paragraph and see at what average x participants can no longer make a "corrected" copy.)

Studies of this form have been done at least on the edge case where all the material removed is from the end (ie. tests of the ability of subjects to predict the next letter in an English text). I'd be interested to see your more general test but am not sure if it has been done. (Except, perhaps, as a game show).