Thought-experimentally: could we potentially be able to scan message databases for the absence of certain phrases, using something like [1], but in a probabilistic manner akin to that of a Bloom filter? This would ensure that law enforcement would be able to flag certain keywords with a nonzero (and nontrivial) false-positive rate. That way, repeated flags end up identifying potentially interesting members of society, without proof and with data inadmissible as reliable evidence in a court of law.
Of course, one runs the risk of the existence of false positives being forgotten, TLA/government pressure to reduce the false positive rate, and so on. But I think this is a slightly interesting way to (partially) preserve privacy while satisfying lawmakers who demand that there be some way for them to listen in on (what should ideally be completely private) data. (This is, of course, only possible once one drops the axiom of privacy being an absolute right: I don't personally support doing this at all.)