Machine learning just on text message metadata can probably be surprisingly accurate on classifying relationships and situations like friend, lover, employee, breakup, drug dealer, prostitute.
Compare
https://www.theguardian.com/technology/2014/mar/13/phone-cal...
which included some manual analysis and use of telephone directories and some automated analysis. (Their data set was tiny and derived from opt-in contributions.)
http://webpolicy.org/2014/03/12/metaphone-the-sensitivity-of...
I'm still hoping to see, and maybe help bring about, better technical and legal protections for communications in my lifetime, or at least better technical protections.