I'm fairly confident that those agencies use context in an automated manner to get any meaningful results.

So "keyword" (could be a word, domain or some other pattern) X may trigger only if Y and Z was already triggered. And some keyword A may only trigger if B was NOT present.

This way you can distinguish doctors, reporters or people studying history or chemistry from those who plan something.

Or e.g. ML applied to patterns over time. Globally.

And yes I do not like it at all, HN is full of people that may likely research some kind of bomb, religion or whatever else out of pure curiosity, but since there are not many such people it can be problem in court one day.

Mix in some Snowden, your hardware stack, gag orders and the fact that we have more laws that anybody can read and you may feel like watching some stupid memes.

OPSEC is about lowering the probability of things going sideways, there are no guarantees either way