I wasn't sure how to write a title that was concise and captured the substance of the article.
It seems like you want to split hairs about whether "analyzing posts" means examining text or metadata. Do you think that "NSA is analyzing phone calls" would be an accurate description of their metadata dragnet?
At any rate, if they aren't looking at the content of self-censored posts yet, it certainly sounds like they're at least interested in doing so. According to the article, "This implies that Facebook wants to know what you are typing in order to understand it."
Normally, the answer to your question is "nothing". In this case -- a bit of short term research -- you should read what the paper actually says:
Content was then marked as “censored” if it was not shared
within the subsequent ten minutes; using this threshold allowed us to record only the presence or absence of text entered, not the keystrokes or content.
If content entered were to go unposted for ten minutes and then be posted, we argue that it was indeed censored (albeit temporarily).
These analyses were conducted in an anonymous manner, so researchers were not privy to any specific user’s activity.
Furthermore, all instrumentation was done on the client side. In other words, the content of self-censored posts and comments was not sent back to Facebook’s servers: Only a binary value that content was entered at all.
It seems like you want to split hairs about whether "analyzing posts" means examining text or metadata. Do you think that "NSA is analyzing phone calls" would be an accurate description of their metadata dragnet?
At any rate, if they aren't looking at the content of self-censored posts yet, it certainly sounds like they're at least interested in doing so. According to the article, "This implies that Facebook wants to know what you are typing in order to understand it."