Also read the first comment:

"Correction: this is a pretty good way to do it. There’s still some big information leaks; [...] For example, a letter responding to a job applicant is likely to be shorter if they got rejected, and longer if they’re hired. And you can do much lower-level analysis, [e.g. sshow-like or keystroke timing attacks]. [...]

Remember that doing bulk statistical analysis of incremental ciphertext edits to gleam as much information as possible is the kind of application that Google’s infrastructure is perfectly designed for!" -- Jim