I am almost positive this has come up before, and I'm equally confident bfv had something useful and insightful to say. Lemme take a crack at it.
What is your profile of Dzhokar Tsarniev? Immigrant, loner, angry, access to weapons, says nasty shit about the US to his friends on the phone, has traveled abroad, possibly to scary places? On the one hand, that covers waaaaay too many people to make it useful. On the other hand, if you decide to push everybody that might be in that pocket, you're likely to radicalize more than a few of them, alienate the rest, and show your hand to anybody you might actually be after. I mean, the CIA is adamant that they didn't radicalize Anwar al-Awlaki. But they totally radicalized Anwar al-Awlaki. Assume one person in a hundred actually intends to do violence. Push a hundred of them. How many intend to do violence now? And this is the NSA we're talking about - they don't push. They have no boots-on-ground. They have wiretaps.
So back it out the other way. List your favorite 30 terrorists and build a profile that matches all of them. What you discover is that (A) if you build it wide enough to include all of them you've included much too large a sample size (B) if you take your sample size down to something reasonable you've excluded a lot of your hits (C) (and C is the real problem) what makes them terrorists generally isn't the shit they say on the phone.
I mean, look.
Joe Lipari might walk into an Apple store on Fifth Avenue with an Armalite AR-10 gas powered semi-automatic weapon and pump round after round into one of those smug, fruity little concierges.
That's a credible threat, method, means and opportunity, from a former Army marksman against a high-profile target. It's also a stand-up comic paraphrasing Fight Club on Facebook. The NYPD took it seriously. The NSA can't because the amount of people suggesting violence in private conversations is orders and orders of magnitude more than the amount of people saying shit in public. It's a real problem: the FBI investigated Omar Mateen for ten fucking months. That's real human beings following a suspect around in unmarked cars'n'shit for longer than a school year and determining the dude wasn't a threat. You're going to pick this guy out via his text messages?
It makes more sense when you realize what the NSA is designed for: spying on zaibatsus and nation-states. Your "find me a terrorist" profile goes to shit but your "find me all communications out of Sary-Shagan related to lasers" or "find me every Siemens contractor that has ever mentioned semiconductors" is super easy. Not only that you can run all of those intercepts together and find commonality. You can get common trends and fill in the blanks. You can do all that old-school Cold War intelligence-gathering goodness that built the whole artifice in the first place.
The link I'm looking for explains how taking a large statistical analysis and backing it out to individual people will never work because the stochastic variation of individuals thwarts any attempt at statistical profiling. However, a large statistical analysis is exactly what you need to determine the size and shape of a nuclear missile program, or clandestine troop movements, or a KGB black site, or a new Samsung semiconductor. And while the first, second and third examples have dropped in priority since the fall of the Iron Curtain, the fourth example is going strong.